1 | //===----------------------------------------------------------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | |
9 | // Copyright (c) Microsoft Corporation. |
10 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
11 | |
12 | // Copyright 2018 Ulf Adams |
13 | // Copyright (c) Microsoft Corporation. All rights reserved. |
14 | |
15 | // Boost Software License - Version 1.0 - August 17th, 2003 |
16 | |
17 | // Permission is hereby granted, free of charge, to any person or organization |
18 | // obtaining a copy of the software and accompanying documentation covered by |
19 | // this license (the "Software") to use, reproduce, display, distribute, |
20 | // execute, and transmit the Software, and to prepare derivative works of the |
21 | // Software, and to permit third-parties to whom the Software is furnished to |
22 | // do so, all subject to the following: |
23 | |
24 | // The copyright notices in the Software and this entire statement, including |
25 | // the above license grant, this restriction and the following disclaimer, |
26 | // must be included in all copies of the Software, in whole or in part, and |
27 | // all derivative works of the Software, unless such copies or derivative |
28 | // works are solely in the form of machine-executable object code generated by |
29 | // a source language processor. |
30 | |
31 | // THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
32 | // IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
33 | // FITNESS FOR A PARTICULAR PURPOSE, TITLE AND NON-INFRINGEMENT. IN NO EVENT |
34 | // SHALL THE COPYRIGHT HOLDERS OR ANYONE DISTRIBUTING THE SOFTWARE BE LIABLE |
35 | // FOR ANY DAMAGES OR OTHER LIABILITY, WHETHER IN CONTRACT, TORT OR OTHERWISE, |
36 | // ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER |
37 | // DEALINGS IN THE SOFTWARE. |
38 | |
39 | // Avoid formatting to keep the changes with the original code minimal. |
40 | // clang-format off |
41 | |
42 | #include <__assert> |
43 | #include <__config> |
44 | #include <charconv> |
45 | #include <cstddef> |
46 | |
47 | #include "include/ryu/common.h" |
48 | #include "include/ryu/d2fixed.h" |
49 | #include "include/ryu/d2s.h" |
50 | #include "include/ryu/d2s_full_table.h" |
51 | #include "include/ryu/d2s_intrinsics.h" |
52 | #include "include/ryu/digit_table.h" |
53 | #include "include/ryu/ryu.h" |
54 | |
55 | _LIBCPP_BEGIN_NAMESPACE_STD |
56 | |
57 | // We need a 64x128-bit multiplication and a subsequent 128-bit shift. |
58 | // Multiplication: |
59 | // The 64-bit factor is variable and passed in, the 128-bit factor comes |
60 | // from a lookup table. We know that the 64-bit factor only has 55 |
61 | // significant bits (i.e., the 9 topmost bits are zeros). The 128-bit |
62 | // factor only has 124 significant bits (i.e., the 4 topmost bits are |
63 | // zeros). |
64 | // Shift: |
65 | // In principle, the multiplication result requires 55 + 124 = 179 bits to |
66 | // represent. However, we then shift this value to the right by __j, which is |
67 | // at least __j >= 115, so the result is guaranteed to fit into 179 - 115 = 64 |
68 | // bits. This means that we only need the topmost 64 significant bits of |
69 | // the 64x128-bit multiplication. |
70 | // |
71 | // There are several ways to do this: |
72 | // 1. Best case: the compiler exposes a 128-bit type. |
73 | // We perform two 64x64-bit multiplications, add the higher 64 bits of the |
74 | // lower result to the higher result, and shift by __j - 64 bits. |
75 | // |
76 | // We explicitly cast from 64-bit to 128-bit, so the compiler can tell |
77 | // that these are only 64-bit inputs, and can map these to the best |
78 | // possible sequence of assembly instructions. |
79 | // x64 machines happen to have matching assembly instructions for |
80 | // 64x64-bit multiplications and 128-bit shifts. |
81 | // |
82 | // 2. Second best case: the compiler exposes intrinsics for the x64 assembly |
83 | // instructions mentioned in 1. |
84 | // |
85 | // 3. We only have 64x64 bit instructions that return the lower 64 bits of |
86 | // the result, i.e., we have to use plain C. |
87 | // Our inputs are less than the full width, so we have three options: |
88 | // a. Ignore this fact and just implement the intrinsics manually. |
89 | // b. Split both into 31-bit pieces, which guarantees no internal overflow, |
90 | // but requires extra work upfront (unless we change the lookup table). |
91 | // c. Split only the first factor into 31-bit pieces, which also guarantees |
92 | // no internal overflow, but requires extra work since the intermediate |
93 | // results are not perfectly aligned. |
94 | #ifdef _LIBCPP_INTRINSIC128 |
95 | |
96 | [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline uint64_t __mulShift(const uint64_t __m, const uint64_t* const __mul, const int32_t __j) { |
97 | // __m is maximum 55 bits |
98 | uint64_t __high1; // 128 |
99 | const uint64_t __low1 = __ryu_umul128(__m, __mul[1], &__high1); // 64 |
100 | uint64_t __high0; // 64 |
101 | (void) __ryu_umul128(__m, __mul[0], &__high0); // 0 |
102 | const uint64_t __sum = __high0 + __low1; |
103 | if (__sum < __high0) { |
104 | ++__high1; // overflow into __high1 |
105 | } |
106 | return __ryu_shiftright128(__sum, __high1, static_cast<uint32_t>(__j - 64)); |
107 | } |
108 | |
109 | [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline uint64_t __mulShiftAll(const uint64_t __m, const uint64_t* const __mul, const int32_t __j, |
110 | uint64_t* const __vp, uint64_t* const __vm, const uint32_t __mmShift) { |
111 | *__vp = __mulShift(4 * __m + 2, __mul, __j); |
112 | *__vm = __mulShift(4 * __m - 1 - __mmShift, __mul, __j); |
113 | return __mulShift(4 * __m, __mul, __j); |
114 | } |
115 | |
116 | #else // ^^^ intrinsics available ^^^ / vvv intrinsics unavailable vvv |
117 | |
118 | [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline _LIBCPP_ALWAYS_INLINE uint64_t __mulShiftAll(uint64_t __m, const uint64_t* const __mul, const int32_t __j, |
119 | uint64_t* const __vp, uint64_t* const __vm, const uint32_t __mmShift) { // TRANSITION, VSO-634761 |
120 | __m <<= 1; |
121 | // __m is maximum 55 bits |
122 | uint64_t __tmp; |
123 | const uint64_t __lo = __ryu_umul128(__m, __mul[0], &__tmp); |
124 | uint64_t __hi; |
125 | const uint64_t __mid = __tmp + __ryu_umul128(__m, __mul[1], &__hi); |
126 | __hi += __mid < __tmp; // overflow into __hi |
127 | |
128 | const uint64_t __lo2 = __lo + __mul[0]; |
129 | const uint64_t __mid2 = __mid + __mul[1] + (__lo2 < __lo); |
130 | const uint64_t __hi2 = __hi + (__mid2 < __mid); |
131 | *__vp = __ryu_shiftright128(__mid2, __hi2, static_cast<uint32_t>(__j - 64 - 1)); |
132 | |
133 | if (__mmShift == 1) { |
134 | const uint64_t __lo3 = __lo - __mul[0]; |
135 | const uint64_t __mid3 = __mid - __mul[1] - (__lo3 > __lo); |
136 | const uint64_t __hi3 = __hi - (__mid3 > __mid); |
137 | *__vm = __ryu_shiftright128(__mid3, __hi3, static_cast<uint32_t>(__j - 64 - 1)); |
138 | } else { |
139 | const uint64_t __lo3 = __lo + __lo; |
140 | const uint64_t __mid3 = __mid + __mid + (__lo3 < __lo); |
141 | const uint64_t __hi3 = __hi + __hi + (__mid3 < __mid); |
142 | const uint64_t __lo4 = __lo3 - __mul[0]; |
143 | const uint64_t __mid4 = __mid3 - __mul[1] - (__lo4 > __lo3); |
144 | const uint64_t __hi4 = __hi3 - (__mid4 > __mid3); |
145 | *__vm = __ryu_shiftright128(__mid4, __hi4, static_cast<uint32_t>(__j - 64)); |
146 | } |
147 | |
148 | return __ryu_shiftright128(__mid, __hi, static_cast<uint32_t>(__j - 64 - 1)); |
149 | } |
150 | |
151 | #endif // ^^^ intrinsics unavailable ^^^ |
152 | |
153 | [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline uint32_t __decimalLength17(const uint64_t __v) { |
154 | // This is slightly faster than a loop. |
155 | // The average output length is 16.38 digits, so we check high-to-low. |
156 | // Function precondition: __v is not an 18, 19, or 20-digit number. |
157 | // (17 digits are sufficient for round-tripping.) |
158 | _LIBCPP_ASSERT_INTERNAL(__v < 100000000000000000u, "" ); |
159 | if (__v >= 10000000000000000u) { return 17; } |
160 | if (__v >= 1000000000000000u) { return 16; } |
161 | if (__v >= 100000000000000u) { return 15; } |
162 | if (__v >= 10000000000000u) { return 14; } |
163 | if (__v >= 1000000000000u) { return 13; } |
164 | if (__v >= 100000000000u) { return 12; } |
165 | if (__v >= 10000000000u) { return 11; } |
166 | if (__v >= 1000000000u) { return 10; } |
167 | if (__v >= 100000000u) { return 9; } |
168 | if (__v >= 10000000u) { return 8; } |
169 | if (__v >= 1000000u) { return 7; } |
170 | if (__v >= 100000u) { return 6; } |
171 | if (__v >= 10000u) { return 5; } |
172 | if (__v >= 1000u) { return 4; } |
173 | if (__v >= 100u) { return 3; } |
174 | if (__v >= 10u) { return 2; } |
175 | return 1; |
176 | } |
177 | |
178 | // A floating decimal representing m * 10^e. |
179 | struct __floating_decimal_64 { |
180 | uint64_t __mantissa; |
181 | int32_t __exponent; |
182 | }; |
183 | |
184 | [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline __floating_decimal_64 __d2d(const uint64_t __ieeeMantissa, const uint32_t __ieeeExponent) { |
185 | int32_t __e2; |
186 | uint64_t __m2; |
187 | if (__ieeeExponent == 0) { |
188 | // We subtract 2 so that the bounds computation has 2 additional bits. |
189 | __e2 = 1 - __DOUBLE_BIAS - __DOUBLE_MANTISSA_BITS - 2; |
190 | __m2 = __ieeeMantissa; |
191 | } else { |
192 | __e2 = static_cast<int32_t>(__ieeeExponent) - __DOUBLE_BIAS - __DOUBLE_MANTISSA_BITS - 2; |
193 | __m2 = (1ull << __DOUBLE_MANTISSA_BITS) | __ieeeMantissa; |
194 | } |
195 | const bool __even = (__m2 & 1) == 0; |
196 | const bool __acceptBounds = __even; |
197 | |
198 | // Step 2: Determine the interval of valid decimal representations. |
199 | const uint64_t __mv = 4 * __m2; |
200 | // Implicit bool -> int conversion. True is 1, false is 0. |
201 | const uint32_t __mmShift = __ieeeMantissa != 0 || __ieeeExponent <= 1; |
202 | // We would compute __mp and __mm like this: |
203 | // uint64_t __mp = 4 * __m2 + 2; |
204 | // uint64_t __mm = __mv - 1 - __mmShift; |
205 | |
206 | // Step 3: Convert to a decimal power base using 128-bit arithmetic. |
207 | uint64_t __vr, __vp, __vm; |
208 | int32_t __e10; |
209 | bool __vmIsTrailingZeros = false; |
210 | bool __vrIsTrailingZeros = false; |
211 | if (__e2 >= 0) { |
212 | // I tried special-casing __q == 0, but there was no effect on performance. |
213 | // This expression is slightly faster than max(0, __log10Pow2(__e2) - 1). |
214 | const uint32_t __q = __log10Pow2(__e2) - (__e2 > 3); |
215 | __e10 = static_cast<int32_t>(__q); |
216 | const int32_t __k = __DOUBLE_POW5_INV_BITCOUNT + __pow5bits(static_cast<int32_t>(__q)) - 1; |
217 | const int32_t __i = -__e2 + static_cast<int32_t>(__q) + __k; |
218 | __vr = __mulShiftAll(__m2, __DOUBLE_POW5_INV_SPLIT[__q], __i, &__vp, &__vm, __mmShift); |
219 | if (__q <= 21) { |
220 | // This should use __q <= 22, but I think 21 is also safe. Smaller values |
221 | // may still be safe, but it's more difficult to reason about them. |
222 | // Only one of __mp, __mv, and __mm can be a multiple of 5, if any. |
223 | const uint32_t __mvMod5 = static_cast<uint32_t>(__mv) - 5 * static_cast<uint32_t>(__div5(__mv)); |
224 | if (__mvMod5 == 0) { |
225 | __vrIsTrailingZeros = __multipleOfPowerOf5(__mv, __q); |
226 | } else if (__acceptBounds) { |
227 | // Same as min(__e2 + (~__mm & 1), __pow5Factor(__mm)) >= __q |
228 | // <=> __e2 + (~__mm & 1) >= __q && __pow5Factor(__mm) >= __q |
229 | // <=> true && __pow5Factor(__mm) >= __q, since __e2 >= __q. |
230 | __vmIsTrailingZeros = __multipleOfPowerOf5(__mv - 1 - __mmShift, __q); |
231 | } else { |
232 | // Same as min(__e2 + 1, __pow5Factor(__mp)) >= __q. |
233 | __vp -= __multipleOfPowerOf5(__mv + 2, __q); |
234 | } |
235 | } |
236 | } else { |
237 | // This expression is slightly faster than max(0, __log10Pow5(-__e2) - 1). |
238 | const uint32_t __q = __log10Pow5(-__e2) - (-__e2 > 1); |
239 | __e10 = static_cast<int32_t>(__q) + __e2; |
240 | const int32_t __i = -__e2 - static_cast<int32_t>(__q); |
241 | const int32_t __k = __pow5bits(__i) - __DOUBLE_POW5_BITCOUNT; |
242 | const int32_t __j = static_cast<int32_t>(__q) - __k; |
243 | __vr = __mulShiftAll(__m2, __DOUBLE_POW5_SPLIT[__i], __j, &__vp, &__vm, __mmShift); |
244 | if (__q <= 1) { |
245 | // {__vr,__vp,__vm} is trailing zeros if {__mv,__mp,__mm} has at least __q trailing 0 bits. |
246 | // __mv = 4 * __m2, so it always has at least two trailing 0 bits. |
247 | __vrIsTrailingZeros = true; |
248 | if (__acceptBounds) { |
249 | // __mm = __mv - 1 - __mmShift, so it has 1 trailing 0 bit iff __mmShift == 1. |
250 | __vmIsTrailingZeros = __mmShift == 1; |
251 | } else { |
252 | // __mp = __mv + 2, so it always has at least one trailing 0 bit. |
253 | --__vp; |
254 | } |
255 | } else if (__q < 63) { // TRANSITION(ulfjack): Use a tighter bound here. |
256 | // We need to compute min(ntz(__mv), __pow5Factor(__mv) - __e2) >= __q - 1 |
257 | // <=> ntz(__mv) >= __q - 1 && __pow5Factor(__mv) - __e2 >= __q - 1 |
258 | // <=> ntz(__mv) >= __q - 1 (__e2 is negative and -__e2 >= __q) |
259 | // <=> (__mv & ((1 << (__q - 1)) - 1)) == 0 |
260 | // We also need to make sure that the left shift does not overflow. |
261 | __vrIsTrailingZeros = __multipleOfPowerOf2(__mv, __q - 1); |
262 | } |
263 | } |
264 | |
265 | // Step 4: Find the shortest decimal representation in the interval of valid representations. |
266 | int32_t __removed = 0; |
267 | uint8_t __lastRemovedDigit = 0; |
268 | uint64_t _Output; |
269 | // On average, we remove ~2 digits. |
270 | if (__vmIsTrailingZeros || __vrIsTrailingZeros) { |
271 | // General case, which happens rarely (~0.7%). |
272 | for (;;) { |
273 | const uint64_t __vpDiv10 = __div10(__vp); |
274 | const uint64_t __vmDiv10 = __div10(__vm); |
275 | if (__vpDiv10 <= __vmDiv10) { |
276 | break; |
277 | } |
278 | const uint32_t __vmMod10 = static_cast<uint32_t>(__vm) - 10 * static_cast<uint32_t>(__vmDiv10); |
279 | const uint64_t __vrDiv10 = __div10(__vr); |
280 | const uint32_t __vrMod10 = static_cast<uint32_t>(__vr) - 10 * static_cast<uint32_t>(__vrDiv10); |
281 | __vmIsTrailingZeros &= __vmMod10 == 0; |
282 | __vrIsTrailingZeros &= __lastRemovedDigit == 0; |
283 | __lastRemovedDigit = static_cast<uint8_t>(__vrMod10); |
284 | __vr = __vrDiv10; |
285 | __vp = __vpDiv10; |
286 | __vm = __vmDiv10; |
287 | ++__removed; |
288 | } |
289 | if (__vmIsTrailingZeros) { |
290 | for (;;) { |
291 | const uint64_t __vmDiv10 = __div10(__vm); |
292 | const uint32_t __vmMod10 = static_cast<uint32_t>(__vm) - 10 * static_cast<uint32_t>(__vmDiv10); |
293 | if (__vmMod10 != 0) { |
294 | break; |
295 | } |
296 | const uint64_t __vpDiv10 = __div10(__vp); |
297 | const uint64_t __vrDiv10 = __div10(__vr); |
298 | const uint32_t __vrMod10 = static_cast<uint32_t>(__vr) - 10 * static_cast<uint32_t>(__vrDiv10); |
299 | __vrIsTrailingZeros &= __lastRemovedDigit == 0; |
300 | __lastRemovedDigit = static_cast<uint8_t>(__vrMod10); |
301 | __vr = __vrDiv10; |
302 | __vp = __vpDiv10; |
303 | __vm = __vmDiv10; |
304 | ++__removed; |
305 | } |
306 | } |
307 | if (__vrIsTrailingZeros && __lastRemovedDigit == 5 && __vr % 2 == 0) { |
308 | // Round even if the exact number is .....50..0. |
309 | __lastRemovedDigit = 4; |
310 | } |
311 | // We need to take __vr + 1 if __vr is outside bounds or we need to round up. |
312 | _Output = __vr + ((__vr == __vm && (!__acceptBounds || !__vmIsTrailingZeros)) || __lastRemovedDigit >= 5); |
313 | } else { |
314 | // Specialized for the common case (~99.3%). Percentages below are relative to this. |
315 | bool __roundUp = false; |
316 | const uint64_t __vpDiv100 = __div100(__vp); |
317 | const uint64_t __vmDiv100 = __div100(__vm); |
318 | if (__vpDiv100 > __vmDiv100) { // Optimization: remove two digits at a time (~86.2%). |
319 | const uint64_t __vrDiv100 = __div100(__vr); |
320 | const uint32_t __vrMod100 = static_cast<uint32_t>(__vr) - 100 * static_cast<uint32_t>(__vrDiv100); |
321 | __roundUp = __vrMod100 >= 50; |
322 | __vr = __vrDiv100; |
323 | __vp = __vpDiv100; |
324 | __vm = __vmDiv100; |
325 | __removed += 2; |
326 | } |
327 | // Loop iterations below (approximately), without optimization above: |
328 | // 0: 0.03%, 1: 13.8%, 2: 70.6%, 3: 14.0%, 4: 1.40%, 5: 0.14%, 6+: 0.02% |
329 | // Loop iterations below (approximately), with optimization above: |
330 | // 0: 70.6%, 1: 27.8%, 2: 1.40%, 3: 0.14%, 4+: 0.02% |
331 | for (;;) { |
332 | const uint64_t __vpDiv10 = __div10(__vp); |
333 | const uint64_t __vmDiv10 = __div10(__vm); |
334 | if (__vpDiv10 <= __vmDiv10) { |
335 | break; |
336 | } |
337 | const uint64_t __vrDiv10 = __div10(__vr); |
338 | const uint32_t __vrMod10 = static_cast<uint32_t>(__vr) - 10 * static_cast<uint32_t>(__vrDiv10); |
339 | __roundUp = __vrMod10 >= 5; |
340 | __vr = __vrDiv10; |
341 | __vp = __vpDiv10; |
342 | __vm = __vmDiv10; |
343 | ++__removed; |
344 | } |
345 | // We need to take __vr + 1 if __vr is outside bounds or we need to round up. |
346 | _Output = __vr + (__vr == __vm || __roundUp); |
347 | } |
348 | const int32_t __exp = __e10 + __removed; |
349 | |
350 | __floating_decimal_64 __fd; |
351 | __fd.__exponent = __exp; |
352 | __fd.__mantissa = _Output; |
353 | return __fd; |
354 | } |
355 | |
356 | [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline to_chars_result __to_chars(char* const _First, char* const _Last, const __floating_decimal_64 __v, |
357 | chars_format _Fmt, const double __f) { |
358 | // Step 5: Print the decimal representation. |
359 | uint64_t _Output = __v.__mantissa; |
360 | int32_t _Ryu_exponent = __v.__exponent; |
361 | const uint32_t __olength = __decimalLength17(_Output); |
362 | int32_t _Scientific_exponent = _Ryu_exponent + static_cast<int32_t>(__olength) - 1; |
363 | |
364 | if (_Fmt == chars_format{}) { |
365 | int32_t _Lower; |
366 | int32_t _Upper; |
367 | |
368 | if (__olength == 1) { |
369 | // Value | Fixed | Scientific |
370 | // 1e-3 | "0.001" | "1e-03" |
371 | // 1e4 | "10000" | "1e+04" |
372 | _Lower = -3; |
373 | _Upper = 4; |
374 | } else { |
375 | // Value | Fixed | Scientific |
376 | // 1234e-7 | "0.0001234" | "1.234e-04" |
377 | // 1234e5 | "123400000" | "1.234e+08" |
378 | _Lower = -static_cast<int32_t>(__olength + 3); |
379 | _Upper = 5; |
380 | } |
381 | |
382 | if (_Lower <= _Ryu_exponent && _Ryu_exponent <= _Upper) { |
383 | _Fmt = chars_format::fixed; |
384 | } else { |
385 | _Fmt = chars_format::scientific; |
386 | } |
387 | } else if (_Fmt == chars_format::general) { |
388 | // C11 7.21.6.1 "The fprintf function"/8: |
389 | // "Let P equal [...] 6 if the precision is omitted [...]. |
390 | // Then, if a conversion with style E would have an exponent of X: |
391 | // - if P > X >= -4, the conversion is with style f [...]. |
392 | // - otherwise, the conversion is with style e [...]." |
393 | if (-4 <= _Scientific_exponent && _Scientific_exponent < 6) { |
394 | _Fmt = chars_format::fixed; |
395 | } else { |
396 | _Fmt = chars_format::scientific; |
397 | } |
398 | } |
399 | |
400 | if (_Fmt == chars_format::fixed) { |
401 | // Example: _Output == 1729, __olength == 4 |
402 | |
403 | // _Ryu_exponent | Printed | _Whole_digits | _Total_fixed_length | Notes |
404 | // --------------|----------|---------------|----------------------|--------------------------------------- |
405 | // 2 | 172900 | 6 | _Whole_digits | Ryu can't be used for printing |
406 | // 1 | 17290 | 5 | (sometimes adjusted) | when the trimmed digits are nonzero. |
407 | // --------------|----------|---------------|----------------------|--------------------------------------- |
408 | // 0 | 1729 | 4 | _Whole_digits | Unified length cases. |
409 | // --------------|----------|---------------|----------------------|--------------------------------------- |
410 | // -1 | 172.9 | 3 | __olength + 1 | This case can't happen for |
411 | // -2 | 17.29 | 2 | | __olength == 1, but no additional |
412 | // -3 | 1.729 | 1 | | code is needed to avoid it. |
413 | // --------------|----------|---------------|----------------------|--------------------------------------- |
414 | // -4 | 0.1729 | 0 | 2 - _Ryu_exponent | C11 7.21.6.1 "The fprintf function"/8: |
415 | // -5 | 0.01729 | -1 | | "If a decimal-point character appears, |
416 | // -6 | 0.001729 | -2 | | at least one digit appears before it." |
417 | |
418 | const int32_t _Whole_digits = static_cast<int32_t>(__olength) + _Ryu_exponent; |
419 | |
420 | uint32_t _Total_fixed_length; |
421 | if (_Ryu_exponent >= 0) { // cases "172900" and "1729" |
422 | _Total_fixed_length = static_cast<uint32_t>(_Whole_digits); |
423 | if (_Output == 1) { |
424 | // Rounding can affect the number of digits. |
425 | // For example, 1e23 is exactly "99999999999999991611392" which is 23 digits instead of 24. |
426 | // We can use a lookup table to detect this and adjust the total length. |
427 | static constexpr uint8_t _Adjustment[309] = { |
428 | 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,1,0,0,0,1,1,0,1,0,1,1,1,0,1,1,1,0,0,0,0,0, |
429 | 1,1,0,0,1,0,1,1,1,0,0,0,0,1,1,1,1,0,0,0,1,1,1,1,0,0,0,1,1,1,1,0,1,0,1,0,1,1,0,0,0,0,1,1,1, |
430 | 1,0,0,0,0,0,0,0,1,1,0,1,1,0,0,1,0,1,0,1,0,1,1,0,0,0,0,0,1,1,1,0,0,1,1,1,1,1,0,1,0,1,1,0,1, |
431 | 1,0,0,0,0,0,0,0,0,0,1,1,1,0,0,1,0,0,1,0,0,1,1,1,1,0,0,1,1,0,1,1,0,1,1,0,1,0,0,0,1,0,0,0,1, |
432 | 0,1,0,1,0,1,1,1,0,0,0,0,0,0,1,1,1,1,0,0,1,0,1,1,1,0,0,0,1,0,1,1,1,1,1,1,0,1,0,1,1,0,0,0,1, |
433 | 1,1,0,1,1,0,0,0,1,0,0,0,1,0,1,0,0,0,0,0,0,0,1,0,1,1,0,0,1,1,1,0,0,0,1,0,1,0,0,0,0,0,1,1,0, |
434 | 0,1,0,1,1,1,0,0,1,0,0,0,0,1,0,1,0,0,0,0,0,1,0,1,0,1,1,0,1,0,0,0,0,0,1,1,0,1,0 }; |
435 | _Total_fixed_length -= _Adjustment[_Ryu_exponent]; |
436 | // _Whole_digits doesn't need to be adjusted because these cases won't refer to it later. |
437 | } |
438 | } else if (_Whole_digits > 0) { // case "17.29" |
439 | _Total_fixed_length = __olength + 1; |
440 | } else { // case "0.001729" |
441 | _Total_fixed_length = static_cast<uint32_t>(2 - _Ryu_exponent); |
442 | } |
443 | |
444 | if (_Last - _First < static_cast<ptrdiff_t>(_Total_fixed_length)) { |
445 | return { _Last, errc::value_too_large }; |
446 | } |
447 | |
448 | char* _Mid; |
449 | if (_Ryu_exponent > 0) { // case "172900" |
450 | bool _Can_use_ryu; |
451 | |
452 | if (_Ryu_exponent > 22) { // 10^22 is the largest power of 10 that's exactly representable as a double. |
453 | _Can_use_ryu = false; |
454 | } else { |
455 | // Ryu generated X: __v.__mantissa * 10^_Ryu_exponent |
456 | // __v.__mantissa == 2^_Trailing_zero_bits * (__v.__mantissa >> _Trailing_zero_bits) |
457 | // 10^_Ryu_exponent == 2^_Ryu_exponent * 5^_Ryu_exponent |
458 | |
459 | // _Trailing_zero_bits is [0, 56] (aside: because 2^56 is the largest power of 2 |
460 | // with 17 decimal digits, which is double's round-trip limit.) |
461 | // _Ryu_exponent is [1, 22]. |
462 | // Normalization adds [2, 52] (aside: at least 2 because the pre-normalized mantissa is at least 5). |
463 | // This adds up to [3, 130], which is well below double's maximum binary exponent 1023. |
464 | |
465 | // Therefore, we just need to consider (__v.__mantissa >> _Trailing_zero_bits) * 5^_Ryu_exponent. |
466 | |
467 | // If that product would exceed 53 bits, then X can't be exactly represented as a double. |
468 | // (That's not a problem for round-tripping, because X is close enough to the original double, |
469 | // but X isn't mathematically equal to the original double.) This requires a high-precision fallback. |
470 | |
471 | // If the product is 53 bits or smaller, then X can be exactly represented as a double (and we don't |
472 | // need to re-synthesize it; the original double must have been X, because Ryu wouldn't produce the |
473 | // same output for two different doubles X and Y). This allows Ryu's output to be used (zero-filled). |
474 | |
475 | // (2^53 - 1) / 5^0 (for indexing), (2^53 - 1) / 5^1, ..., (2^53 - 1) / 5^22 |
476 | static constexpr uint64_t _Max_shifted_mantissa[23] = { |
477 | 9007199254740991u, 1801439850948198u, 360287970189639u, 72057594037927u, 14411518807585u, |
478 | 2882303761517u, 576460752303u, 115292150460u, 23058430092u, 4611686018u, 922337203u, 184467440u, |
479 | 36893488u, 7378697u, 1475739u, 295147u, 59029u, 11805u, 2361u, 472u, 94u, 18u, 3u }; |
480 | |
481 | unsigned long _Trailing_zero_bits; |
482 | #if _LIBCPP_HAS_BITSCAN64 |
483 | (void) _BitScanForward64(&_Trailing_zero_bits, __v.__mantissa); // __v.__mantissa is guaranteed nonzero |
484 | #else // ^^^ 64-bit ^^^ / vvv 32-bit vvv |
485 | const uint32_t _Low_mantissa = static_cast<uint32_t>(__v.__mantissa); |
486 | if (_Low_mantissa != 0) { |
487 | (void) _BitScanForward(&_Trailing_zero_bits, _Low_mantissa); |
488 | } else { |
489 | const uint32_t _High_mantissa = static_cast<uint32_t>(__v.__mantissa >> 32); // nonzero here |
490 | (void) _BitScanForward(&_Trailing_zero_bits, _High_mantissa); |
491 | _Trailing_zero_bits += 32; |
492 | } |
493 | #endif // ^^^ 32-bit ^^^ |
494 | const uint64_t _Shifted_mantissa = __v.__mantissa >> _Trailing_zero_bits; |
495 | _Can_use_ryu = _Shifted_mantissa <= _Max_shifted_mantissa[_Ryu_exponent]; |
496 | } |
497 | |
498 | if (!_Can_use_ryu) { |
499 | // Print the integer exactly. |
500 | // Performance note: This will redundantly perform bounds checking. |
501 | // Performance note: This will redundantly decompose the IEEE representation. |
502 | return __d2fixed_buffered_n(_First, _Last, __f, 0); |
503 | } |
504 | |
505 | // _Can_use_ryu |
506 | // Print the decimal digits, left-aligned within [_First, _First + _Total_fixed_length). |
507 | _Mid = _First + __olength; |
508 | } else { // cases "1729", "17.29", and "0.001729" |
509 | // Print the decimal digits, right-aligned within [_First, _First + _Total_fixed_length). |
510 | _Mid = _First + _Total_fixed_length; |
511 | } |
512 | |
513 | // We prefer 32-bit operations, even on 64-bit platforms. |
514 | // We have at most 17 digits, and uint32_t can store 9 digits. |
515 | // If _Output doesn't fit into uint32_t, we cut off 8 digits, |
516 | // so the rest will fit into uint32_t. |
517 | if ((_Output >> 32) != 0) { |
518 | // Expensive 64-bit division. |
519 | const uint64_t __q = __div1e8(_Output); |
520 | uint32_t __output2 = static_cast<uint32_t>(_Output - 100000000 * __q); |
521 | _Output = __q; |
522 | |
523 | const uint32_t __c = __output2 % 10000; |
524 | __output2 /= 10000; |
525 | const uint32_t __d = __output2 % 10000; |
526 | const uint32_t __c0 = (__c % 100) << 1; |
527 | const uint32_t __c1 = (__c / 100) << 1; |
528 | const uint32_t __d0 = (__d % 100) << 1; |
529 | const uint32_t __d1 = (__d / 100) << 1; |
530 | |
531 | std::memcpy(_Mid -= 2, __DIGIT_TABLE + __c0, 2); |
532 | std::memcpy(_Mid -= 2, __DIGIT_TABLE + __c1, 2); |
533 | std::memcpy(_Mid -= 2, __DIGIT_TABLE + __d0, 2); |
534 | std::memcpy(_Mid -= 2, __DIGIT_TABLE + __d1, 2); |
535 | } |
536 | uint32_t __output2 = static_cast<uint32_t>(_Output); |
537 | while (__output2 >= 10000) { |
538 | #ifdef __clang__ // TRANSITION, LLVM-38217 |
539 | const uint32_t __c = __output2 - 10000 * (__output2 / 10000); |
540 | #else |
541 | const uint32_t __c = __output2 % 10000; |
542 | #endif |
543 | __output2 /= 10000; |
544 | const uint32_t __c0 = (__c % 100) << 1; |
545 | const uint32_t __c1 = (__c / 100) << 1; |
546 | std::memcpy(_Mid -= 2, __DIGIT_TABLE + __c0, 2); |
547 | std::memcpy(_Mid -= 2, __DIGIT_TABLE + __c1, 2); |
548 | } |
549 | if (__output2 >= 100) { |
550 | const uint32_t __c = (__output2 % 100) << 1; |
551 | __output2 /= 100; |
552 | std::memcpy(_Mid -= 2, __DIGIT_TABLE + __c, 2); |
553 | } |
554 | if (__output2 >= 10) { |
555 | const uint32_t __c = __output2 << 1; |
556 | std::memcpy(_Mid -= 2, __DIGIT_TABLE + __c, 2); |
557 | } else { |
558 | *--_Mid = static_cast<char>('0' + __output2); |
559 | } |
560 | |
561 | if (_Ryu_exponent > 0) { // case "172900" with _Can_use_ryu |
562 | // Performance note: it might be more efficient to do this immediately after setting _Mid. |
563 | std::memset(_First + __olength, '0', static_cast<size_t>(_Ryu_exponent)); |
564 | } else if (_Ryu_exponent == 0) { // case "1729" |
565 | // Done! |
566 | } else if (_Whole_digits > 0) { // case "17.29" |
567 | // Performance note: moving digits might not be optimal. |
568 | std::memmove(_First, _First + 1, static_cast<size_t>(_Whole_digits)); |
569 | _First[_Whole_digits] = '.'; |
570 | } else { // case "0.001729" |
571 | // Performance note: a larger memset() followed by overwriting '.' might be more efficient. |
572 | _First[0] = '0'; |
573 | _First[1] = '.'; |
574 | std::memset(_First + 2, '0', static_cast<size_t>(-_Whole_digits)); |
575 | } |
576 | |
577 | return { _First + _Total_fixed_length, errc{} }; |
578 | } |
579 | |
580 | const uint32_t _Total_scientific_length = __olength + (__olength > 1) // digits + possible decimal point |
581 | + (-100 < _Scientific_exponent && _Scientific_exponent < 100 ? 4 : 5); // + scientific exponent |
582 | if (_Last - _First < static_cast<ptrdiff_t>(_Total_scientific_length)) { |
583 | return { _Last, errc::value_too_large }; |
584 | } |
585 | char* const __result = _First; |
586 | |
587 | // Print the decimal digits. |
588 | uint32_t __i = 0; |
589 | // We prefer 32-bit operations, even on 64-bit platforms. |
590 | // We have at most 17 digits, and uint32_t can store 9 digits. |
591 | // If _Output doesn't fit into uint32_t, we cut off 8 digits, |
592 | // so the rest will fit into uint32_t. |
593 | if ((_Output >> 32) != 0) { |
594 | // Expensive 64-bit division. |
595 | const uint64_t __q = __div1e8(_Output); |
596 | uint32_t __output2 = static_cast<uint32_t>(_Output) - 100000000 * static_cast<uint32_t>(__q); |
597 | _Output = __q; |
598 | |
599 | const uint32_t __c = __output2 % 10000; |
600 | __output2 /= 10000; |
601 | const uint32_t __d = __output2 % 10000; |
602 | const uint32_t __c0 = (__c % 100) << 1; |
603 | const uint32_t __c1 = (__c / 100) << 1; |
604 | const uint32_t __d0 = (__d % 100) << 1; |
605 | const uint32_t __d1 = (__d / 100) << 1; |
606 | std::memcpy(__result + __olength - __i - 1, __DIGIT_TABLE + __c0, 2); |
607 | std::memcpy(__result + __olength - __i - 3, __DIGIT_TABLE + __c1, 2); |
608 | std::memcpy(__result + __olength - __i - 5, __DIGIT_TABLE + __d0, 2); |
609 | std::memcpy(__result + __olength - __i - 7, __DIGIT_TABLE + __d1, 2); |
610 | __i += 8; |
611 | } |
612 | uint32_t __output2 = static_cast<uint32_t>(_Output); |
613 | while (__output2 >= 10000) { |
614 | #ifdef __clang__ // TRANSITION, LLVM-38217 |
615 | const uint32_t __c = __output2 - 10000 * (__output2 / 10000); |
616 | #else |
617 | const uint32_t __c = __output2 % 10000; |
618 | #endif |
619 | __output2 /= 10000; |
620 | const uint32_t __c0 = (__c % 100) << 1; |
621 | const uint32_t __c1 = (__c / 100) << 1; |
622 | std::memcpy(__result + __olength - __i - 1, __DIGIT_TABLE + __c0, 2); |
623 | std::memcpy(__result + __olength - __i - 3, __DIGIT_TABLE + __c1, 2); |
624 | __i += 4; |
625 | } |
626 | if (__output2 >= 100) { |
627 | const uint32_t __c = (__output2 % 100) << 1; |
628 | __output2 /= 100; |
629 | std::memcpy(__result + __olength - __i - 1, __DIGIT_TABLE + __c, 2); |
630 | __i += 2; |
631 | } |
632 | if (__output2 >= 10) { |
633 | const uint32_t __c = __output2 << 1; |
634 | // We can't use memcpy here: the decimal dot goes between these two digits. |
635 | __result[2] = __DIGIT_TABLE[__c + 1]; |
636 | __result[0] = __DIGIT_TABLE[__c]; |
637 | } else { |
638 | __result[0] = static_cast<char>('0' + __output2); |
639 | } |
640 | |
641 | // Print decimal point if needed. |
642 | uint32_t __index; |
643 | if (__olength > 1) { |
644 | __result[1] = '.'; |
645 | __index = __olength + 1; |
646 | } else { |
647 | __index = 1; |
648 | } |
649 | |
650 | // Print the exponent. |
651 | __result[__index++] = 'e'; |
652 | if (_Scientific_exponent < 0) { |
653 | __result[__index++] = '-'; |
654 | _Scientific_exponent = -_Scientific_exponent; |
655 | } else { |
656 | __result[__index++] = '+'; |
657 | } |
658 | |
659 | if (_Scientific_exponent >= 100) { |
660 | const int32_t __c = _Scientific_exponent % 10; |
661 | std::memcpy(__result + __index, __DIGIT_TABLE + 2 * (_Scientific_exponent / 10), 2); |
662 | __result[__index + 2] = static_cast<char>('0' + __c); |
663 | __index += 3; |
664 | } else { |
665 | std::memcpy(__result + __index, __DIGIT_TABLE + 2 * _Scientific_exponent, 2); |
666 | __index += 2; |
667 | } |
668 | |
669 | return { _First + _Total_scientific_length, errc{} }; |
670 | } |
671 | |
672 | [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline bool __d2d_small_int(const uint64_t __ieeeMantissa, const uint32_t __ieeeExponent, |
673 | __floating_decimal_64* const __v) { |
674 | const uint64_t __m2 = (1ull << __DOUBLE_MANTISSA_BITS) | __ieeeMantissa; |
675 | const int32_t __e2 = static_cast<int32_t>(__ieeeExponent) - __DOUBLE_BIAS - __DOUBLE_MANTISSA_BITS; |
676 | |
677 | if (__e2 > 0) { |
678 | // f = __m2 * 2^__e2 >= 2^53 is an integer. |
679 | // Ignore this case for now. |
680 | return false; |
681 | } |
682 | |
683 | if (__e2 < -52) { |
684 | // f < 1. |
685 | return false; |
686 | } |
687 | |
688 | // Since 2^52 <= __m2 < 2^53 and 0 <= -__e2 <= 52: 1 <= f = __m2 / 2^-__e2 < 2^53. |
689 | // Test if the lower -__e2 bits of the significand are 0, i.e. whether the fraction is 0. |
690 | const uint64_t __mask = (1ull << -__e2) - 1; |
691 | const uint64_t __fraction = __m2 & __mask; |
692 | if (__fraction != 0) { |
693 | return false; |
694 | } |
695 | |
696 | // f is an integer in the range [1, 2^53). |
697 | // Note: __mantissa might contain trailing (decimal) 0's. |
698 | // Note: since 2^53 < 10^16, there is no need to adjust __decimalLength17(). |
699 | __v->__mantissa = __m2 >> -__e2; |
700 | __v->__exponent = 0; |
701 | return true; |
702 | } |
703 | |
704 | [[nodiscard]] to_chars_result __d2s_buffered_n(char* const _First, char* const _Last, const double __f, |
705 | const chars_format _Fmt) { |
706 | |
707 | // Step 1: Decode the floating-point number, and unify normalized and subnormal cases. |
708 | const uint64_t __bits = __double_to_bits(__f); |
709 | |
710 | // Case distinction; exit early for the easy cases. |
711 | if (__bits == 0) { |
712 | if (_Fmt == chars_format::scientific) { |
713 | if (_Last - _First < 5) { |
714 | return { _Last, errc::value_too_large }; |
715 | } |
716 | |
717 | std::memcpy(_First, "0e+00" , 5); |
718 | |
719 | return { _First + 5, errc{} }; |
720 | } |
721 | |
722 | // Print "0" for chars_format::fixed, chars_format::general, and chars_format{}. |
723 | if (_First == _Last) { |
724 | return { _Last, errc::value_too_large }; |
725 | } |
726 | |
727 | *_First = '0'; |
728 | |
729 | return { _First + 1, errc{} }; |
730 | } |
731 | |
732 | // Decode __bits into mantissa and exponent. |
733 | const uint64_t __ieeeMantissa = __bits & ((1ull << __DOUBLE_MANTISSA_BITS) - 1); |
734 | const uint32_t __ieeeExponent = static_cast<uint32_t>(__bits >> __DOUBLE_MANTISSA_BITS); |
735 | |
736 | if (_Fmt == chars_format::fixed) { |
737 | // const uint64_t _Mantissa2 = __ieeeMantissa | (1ull << __DOUBLE_MANTISSA_BITS); // restore implicit bit |
738 | const int32_t _Exponent2 = static_cast<int32_t>(__ieeeExponent) |
739 | - __DOUBLE_BIAS - __DOUBLE_MANTISSA_BITS; // bias and normalization |
740 | |
741 | // Normal values are equal to _Mantissa2 * 2^_Exponent2. |
742 | // (Subnormals are different, but they'll be rejected by the _Exponent2 test here, so they can be ignored.) |
743 | |
744 | // For nonzero integers, _Exponent2 >= -52. (The minimum value occurs when _Mantissa2 * 2^_Exponent2 is 1. |
745 | // In that case, _Mantissa2 is the implicit 1 bit followed by 52 zeros, so _Exponent2 is -52 to shift away |
746 | // the zeros.) The dense range of exactly representable integers has negative or zero exponents |
747 | // (as positive exponents make the range non-dense). For that dense range, Ryu will always be used: |
748 | // every digit is necessary to uniquely identify the value, so Ryu must print them all. |
749 | |
750 | // Positive exponents are the non-dense range of exactly representable integers. This contains all of the values |
751 | // for which Ryu can't be used (and a few Ryu-friendly values). We can save time by detecting positive |
752 | // exponents here and skipping Ryu. Calling __d2fixed_buffered_n() with precision 0 is valid for all integers |
753 | // (so it's okay if we call it with a Ryu-friendly value). |
754 | if (_Exponent2 > 0) { |
755 | return __d2fixed_buffered_n(_First, _Last, __f, 0); |
756 | } |
757 | } |
758 | |
759 | __floating_decimal_64 __v; |
760 | const bool __isSmallInt = __d2d_small_int(__ieeeMantissa, __ieeeExponent, &__v); |
761 | if (__isSmallInt) { |
762 | // For small integers in the range [1, 2^53), __v.__mantissa might contain trailing (decimal) zeros. |
763 | // For scientific notation we need to move these zeros into the exponent. |
764 | // (This is not needed for fixed-point notation, so it might be beneficial to trim |
765 | // trailing zeros in __to_chars only if needed - once fixed-point notation output is implemented.) |
766 | for (;;) { |
767 | const uint64_t __q = __div10(__v.__mantissa); |
768 | const uint32_t __r = static_cast<uint32_t>(__v.__mantissa) - 10 * static_cast<uint32_t>(__q); |
769 | if (__r != 0) { |
770 | break; |
771 | } |
772 | __v.__mantissa = __q; |
773 | ++__v.__exponent; |
774 | } |
775 | } else { |
776 | __v = __d2d(__ieeeMantissa, __ieeeExponent); |
777 | } |
778 | |
779 | return __to_chars(_First, _Last, __v, _Fmt, __f); |
780 | } |
781 | |
782 | _LIBCPP_END_NAMESPACE_STD |
783 | |
784 | // clang-format on |
785 | |