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 <cstdint> |
46 | #include <cstddef> |
47 | |
48 | #include "include/ryu/common.h" |
49 | #include "include/ryu/d2fixed.h" |
50 | #include "include/ryu/d2s_intrinsics.h" |
51 | #include "include/ryu/digit_table.h" |
52 | #include "include/ryu/f2s.h" |
53 | #include "include/ryu/ryu.h" |
54 | |
55 | _LIBCPP_BEGIN_NAMESPACE_STD |
56 | |
57 | inline constexpr int __FLOAT_MANTISSA_BITS = 23; |
58 | inline constexpr int __FLOAT_EXPONENT_BITS = 8; |
59 | inline constexpr int __FLOAT_BIAS = 127; |
60 | |
61 | inline constexpr int __FLOAT_POW5_INV_BITCOUNT = 59; |
62 | inline constexpr uint64_t __FLOAT_POW5_INV_SPLIT[31] = { |
63 | 576460752303423489u, 461168601842738791u, 368934881474191033u, 295147905179352826u, |
64 | 472236648286964522u, 377789318629571618u, 302231454903657294u, 483570327845851670u, |
65 | 386856262276681336u, 309485009821345069u, 495176015714152110u, 396140812571321688u, |
66 | 316912650057057351u, 507060240091291761u, 405648192073033409u, 324518553658426727u, |
67 | 519229685853482763u, 415383748682786211u, 332306998946228969u, 531691198313966350u, |
68 | 425352958651173080u, 340282366920938464u, 544451787073501542u, 435561429658801234u, |
69 | 348449143727040987u, 557518629963265579u, 446014903970612463u, 356811923176489971u, |
70 | 570899077082383953u, 456719261665907162u, 365375409332725730u |
71 | }; |
72 | inline constexpr int __FLOAT_POW5_BITCOUNT = 61; |
73 | inline constexpr uint64_t __FLOAT_POW5_SPLIT[47] = { |
74 | 1152921504606846976u, 1441151880758558720u, 1801439850948198400u, 2251799813685248000u, |
75 | 1407374883553280000u, 1759218604441600000u, 2199023255552000000u, 1374389534720000000u, |
76 | 1717986918400000000u, 2147483648000000000u, 1342177280000000000u, 1677721600000000000u, |
77 | 2097152000000000000u, 1310720000000000000u, 1638400000000000000u, 2048000000000000000u, |
78 | 1280000000000000000u, 1600000000000000000u, 2000000000000000000u, 1250000000000000000u, |
79 | 1562500000000000000u, 1953125000000000000u, 1220703125000000000u, 1525878906250000000u, |
80 | 1907348632812500000u, 1192092895507812500u, 1490116119384765625u, 1862645149230957031u, |
81 | 1164153218269348144u, 1455191522836685180u, 1818989403545856475u, 2273736754432320594u, |
82 | 1421085471520200371u, 1776356839400250464u, 2220446049250313080u, 1387778780781445675u, |
83 | 1734723475976807094u, 2168404344971008868u, 1355252715606880542u, 1694065894508600678u, |
84 | 2117582368135750847u, 1323488980084844279u, 1654361225106055349u, 2067951531382569187u, |
85 | 1292469707114105741u, 1615587133892632177u, 2019483917365790221u |
86 | }; |
87 | |
88 | [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline uint32_t __pow5Factor(uint32_t __value) { |
89 | uint32_t __count = 0; |
90 | for (;;) { |
91 | _LIBCPP_ASSERT_INTERNAL(__value != 0, "" ); |
92 | const uint32_t __q = __value / 5; |
93 | const uint32_t __r = __value % 5; |
94 | if (__r != 0) { |
95 | break; |
96 | } |
97 | __value = __q; |
98 | ++__count; |
99 | } |
100 | return __count; |
101 | } |
102 | |
103 | // Returns true if __value is divisible by 5^__p. |
104 | [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline bool __multipleOfPowerOf5(const uint32_t __value, const uint32_t __p) { |
105 | return __pow5Factor(__value) >= __p; |
106 | } |
107 | |
108 | // Returns true if __value is divisible by 2^__p. |
109 | [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline bool __multipleOfPowerOf2(const uint32_t __value, const uint32_t __p) { |
110 | _LIBCPP_ASSERT_INTERNAL(__value != 0, "" ); |
111 | _LIBCPP_ASSERT_INTERNAL(__p < 32, "" ); |
112 | // __builtin_ctz doesn't appear to be faster here. |
113 | return (__value & ((1u << __p) - 1)) == 0; |
114 | } |
115 | |
116 | [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline uint32_t __mulShift(const uint32_t __m, const uint64_t __factor, const int32_t __shift) { |
117 | _LIBCPP_ASSERT_INTERNAL(__shift > 32, "" ); |
118 | |
119 | // The casts here help MSVC to avoid calls to the __allmul library |
120 | // function. |
121 | const uint32_t __factorLo = static_cast<uint32_t>(__factor); |
122 | const uint32_t __factorHi = static_cast<uint32_t>(__factor >> 32); |
123 | const uint64_t __bits0 = static_cast<uint64_t>(__m) * __factorLo; |
124 | const uint64_t __bits1 = static_cast<uint64_t>(__m) * __factorHi; |
125 | |
126 | #ifndef _LIBCPP_64_BIT |
127 | // On 32-bit platforms we can avoid a 64-bit shift-right since we only |
128 | // need the upper 32 bits of the result and the shift value is > 32. |
129 | const uint32_t __bits0Hi = static_cast<uint32_t>(__bits0 >> 32); |
130 | uint32_t __bits1Lo = static_cast<uint32_t>(__bits1); |
131 | uint32_t __bits1Hi = static_cast<uint32_t>(__bits1 >> 32); |
132 | __bits1Lo += __bits0Hi; |
133 | __bits1Hi += (__bits1Lo < __bits0Hi); |
134 | const int32_t __s = __shift - 32; |
135 | return (__bits1Hi << (32 - __s)) | (__bits1Lo >> __s); |
136 | #else // ^^^ 32-bit ^^^ / vvv 64-bit vvv |
137 | const uint64_t __sum = (__bits0 >> 32) + __bits1; |
138 | const uint64_t __shiftedSum = __sum >> (__shift - 32); |
139 | _LIBCPP_ASSERT_INTERNAL(__shiftedSum <= UINT32_MAX, "" ); |
140 | return static_cast<uint32_t>(__shiftedSum); |
141 | #endif // ^^^ 64-bit ^^^ |
142 | } |
143 | |
144 | [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline uint32_t __mulPow5InvDivPow2(const uint32_t __m, const uint32_t __q, const int32_t __j) { |
145 | return __mulShift(__m, __FLOAT_POW5_INV_SPLIT[__q], __j); |
146 | } |
147 | |
148 | [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline uint32_t __mulPow5divPow2(const uint32_t __m, const uint32_t __i, const int32_t __j) { |
149 | return __mulShift(__m, __FLOAT_POW5_SPLIT[__i], __j); |
150 | } |
151 | |
152 | // A floating decimal representing m * 10^e. |
153 | struct __floating_decimal_32 { |
154 | uint32_t __mantissa; |
155 | int32_t __exponent; |
156 | }; |
157 | |
158 | [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline __floating_decimal_32 __f2d(const uint32_t __ieeeMantissa, const uint32_t __ieeeExponent) { |
159 | int32_t __e2; |
160 | uint32_t __m2; |
161 | if (__ieeeExponent == 0) { |
162 | // We subtract 2 so that the bounds computation has 2 additional bits. |
163 | __e2 = 1 - __FLOAT_BIAS - __FLOAT_MANTISSA_BITS - 2; |
164 | __m2 = __ieeeMantissa; |
165 | } else { |
166 | __e2 = static_cast<int32_t>(__ieeeExponent) - __FLOAT_BIAS - __FLOAT_MANTISSA_BITS - 2; |
167 | __m2 = (1u << __FLOAT_MANTISSA_BITS) | __ieeeMantissa; |
168 | } |
169 | const bool __even = (__m2 & 1) == 0; |
170 | const bool __acceptBounds = __even; |
171 | |
172 | // Step 2: Determine the interval of valid decimal representations. |
173 | const uint32_t __mv = 4 * __m2; |
174 | const uint32_t __mp = 4 * __m2 + 2; |
175 | // Implicit bool -> int conversion. True is 1, false is 0. |
176 | const uint32_t __mmShift = __ieeeMantissa != 0 || __ieeeExponent <= 1; |
177 | const uint32_t __mm = 4 * __m2 - 1 - __mmShift; |
178 | |
179 | // Step 3: Convert to a decimal power base using 64-bit arithmetic. |
180 | uint32_t __vr, __vp, __vm; |
181 | int32_t __e10; |
182 | bool __vmIsTrailingZeros = false; |
183 | bool __vrIsTrailingZeros = false; |
184 | uint8_t __lastRemovedDigit = 0; |
185 | if (__e2 >= 0) { |
186 | const uint32_t __q = __log10Pow2(__e2); |
187 | __e10 = static_cast<int32_t>(__q); |
188 | const int32_t __k = __FLOAT_POW5_INV_BITCOUNT + __pow5bits(static_cast<int32_t>(__q)) - 1; |
189 | const int32_t __i = -__e2 + static_cast<int32_t>(__q) + __k; |
190 | __vr = __mulPow5InvDivPow2(__mv, __q, __i); |
191 | __vp = __mulPow5InvDivPow2(__mp, __q, __i); |
192 | __vm = __mulPow5InvDivPow2(__mm, __q, __i); |
193 | if (__q != 0 && (__vp - 1) / 10 <= __vm / 10) { |
194 | // We need to know one removed digit even if we are not going to loop below. We could use |
195 | // __q = X - 1 above, except that would require 33 bits for the result, and we've found that |
196 | // 32-bit arithmetic is faster even on 64-bit machines. |
197 | const int32_t __l = __FLOAT_POW5_INV_BITCOUNT + __pow5bits(static_cast<int32_t>(__q - 1)) - 1; |
198 | __lastRemovedDigit = static_cast<uint8_t>(__mulPow5InvDivPow2(__mv, __q - 1, |
199 | -__e2 + static_cast<int32_t>(__q) - 1 + __l) % 10); |
200 | } |
201 | if (__q <= 9) { |
202 | // The largest power of 5 that fits in 24 bits is 5^10, but __q <= 9 seems to be safe as well. |
203 | // Only one of __mp, __mv, and __mm can be a multiple of 5, if any. |
204 | if (__mv % 5 == 0) { |
205 | __vrIsTrailingZeros = __multipleOfPowerOf5(__mv, __q); |
206 | } else if (__acceptBounds) { |
207 | __vmIsTrailingZeros = __multipleOfPowerOf5(__mm, __q); |
208 | } else { |
209 | __vp -= __multipleOfPowerOf5(__mp, __q); |
210 | } |
211 | } |
212 | } else { |
213 | const uint32_t __q = __log10Pow5(-__e2); |
214 | __e10 = static_cast<int32_t>(__q) + __e2; |
215 | const int32_t __i = -__e2 - static_cast<int32_t>(__q); |
216 | const int32_t __k = __pow5bits(__i) - __FLOAT_POW5_BITCOUNT; |
217 | int32_t __j = static_cast<int32_t>(__q) - __k; |
218 | __vr = __mulPow5divPow2(__mv, static_cast<uint32_t>(__i), __j); |
219 | __vp = __mulPow5divPow2(__mp, static_cast<uint32_t>(__i), __j); |
220 | __vm = __mulPow5divPow2(__mm, static_cast<uint32_t>(__i), __j); |
221 | if (__q != 0 && (__vp - 1) / 10 <= __vm / 10) { |
222 | __j = static_cast<int32_t>(__q) - 1 - (__pow5bits(__i + 1) - __FLOAT_POW5_BITCOUNT); |
223 | __lastRemovedDigit = static_cast<uint8_t>(__mulPow5divPow2(__mv, static_cast<uint32_t>(__i + 1), __j) % 10); |
224 | } |
225 | if (__q <= 1) { |
226 | // {__vr,__vp,__vm} is trailing zeros if {__mv,__mp,__mm} has at least __q trailing 0 bits. |
227 | // __mv = 4 * __m2, so it always has at least two trailing 0 bits. |
228 | __vrIsTrailingZeros = true; |
229 | if (__acceptBounds) { |
230 | // __mm = __mv - 1 - __mmShift, so it has 1 trailing 0 bit iff __mmShift == 1. |
231 | __vmIsTrailingZeros = __mmShift == 1; |
232 | } else { |
233 | // __mp = __mv + 2, so it always has at least one trailing 0 bit. |
234 | --__vp; |
235 | } |
236 | } else if (__q < 31) { // TRANSITION(ulfjack): Use a tighter bound here. |
237 | __vrIsTrailingZeros = __multipleOfPowerOf2(__mv, __q - 1); |
238 | } |
239 | } |
240 | |
241 | // Step 4: Find the shortest decimal representation in the interval of valid representations. |
242 | int32_t __removed = 0; |
243 | uint32_t _Output; |
244 | if (__vmIsTrailingZeros || __vrIsTrailingZeros) { |
245 | // General case, which happens rarely (~4.0%). |
246 | while (__vp / 10 > __vm / 10) { |
247 | #ifdef __clang__ // TRANSITION, LLVM-23106 |
248 | __vmIsTrailingZeros &= __vm - (__vm / 10) * 10 == 0; |
249 | #else |
250 | __vmIsTrailingZeros &= __vm % 10 == 0; |
251 | #endif |
252 | __vrIsTrailingZeros &= __lastRemovedDigit == 0; |
253 | __lastRemovedDigit = static_cast<uint8_t>(__vr % 10); |
254 | __vr /= 10; |
255 | __vp /= 10; |
256 | __vm /= 10; |
257 | ++__removed; |
258 | } |
259 | if (__vmIsTrailingZeros) { |
260 | while (__vm % 10 == 0) { |
261 | __vrIsTrailingZeros &= __lastRemovedDigit == 0; |
262 | __lastRemovedDigit = static_cast<uint8_t>(__vr % 10); |
263 | __vr /= 10; |
264 | __vp /= 10; |
265 | __vm /= 10; |
266 | ++__removed; |
267 | } |
268 | } |
269 | if (__vrIsTrailingZeros && __lastRemovedDigit == 5 && __vr % 2 == 0) { |
270 | // Round even if the exact number is .....50..0. |
271 | __lastRemovedDigit = 4; |
272 | } |
273 | // We need to take __vr + 1 if __vr is outside bounds or we need to round up. |
274 | _Output = __vr + ((__vr == __vm && (!__acceptBounds || !__vmIsTrailingZeros)) || __lastRemovedDigit >= 5); |
275 | } else { |
276 | // Specialized for the common case (~96.0%). Percentages below are relative to this. |
277 | // Loop iterations below (approximately): |
278 | // 0: 13.6%, 1: 70.7%, 2: 14.1%, 3: 1.39%, 4: 0.14%, 5+: 0.01% |
279 | while (__vp / 10 > __vm / 10) { |
280 | __lastRemovedDigit = static_cast<uint8_t>(__vr % 10); |
281 | __vr /= 10; |
282 | __vp /= 10; |
283 | __vm /= 10; |
284 | ++__removed; |
285 | } |
286 | // We need to take __vr + 1 if __vr is outside bounds or we need to round up. |
287 | _Output = __vr + (__vr == __vm || __lastRemovedDigit >= 5); |
288 | } |
289 | const int32_t __exp = __e10 + __removed; |
290 | |
291 | __floating_decimal_32 __fd; |
292 | __fd.__exponent = __exp; |
293 | __fd.__mantissa = _Output; |
294 | return __fd; |
295 | } |
296 | |
297 | [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline to_chars_result _Large_integer_to_chars(char* const _First, char* const _Last, |
298 | const uint32_t _Mantissa2, const int32_t _Exponent2) { |
299 | |
300 | // Print the integer _Mantissa2 * 2^_Exponent2 exactly. |
301 | |
302 | // For nonzero integers, _Exponent2 >= -23. (The minimum value occurs when _Mantissa2 * 2^_Exponent2 is 1. |
303 | // In that case, _Mantissa2 is the implicit 1 bit followed by 23 zeros, so _Exponent2 is -23 to shift away |
304 | // the zeros.) The dense range of exactly representable integers has negative or zero exponents |
305 | // (as positive exponents make the range non-dense). For that dense range, Ryu will always be used: |
306 | // every digit is necessary to uniquely identify the value, so Ryu must print them all. |
307 | |
308 | // Positive exponents are the non-dense range of exactly representable integers. |
309 | // This contains all of the values for which Ryu can't be used (and a few Ryu-friendly values). |
310 | |
311 | // Performance note: Long division appears to be faster than losslessly widening float to double and calling |
312 | // __d2fixed_buffered_n(). If __f2fixed_buffered_n() is implemented, it might be faster than long division. |
313 | |
314 | _LIBCPP_ASSERT_INTERNAL(_Exponent2 > 0, "" ); |
315 | _LIBCPP_ASSERT_INTERNAL(_Exponent2 <= 104, "" ); // because __ieeeExponent <= 254 |
316 | |
317 | // Manually represent _Mantissa2 * 2^_Exponent2 as a large integer. _Mantissa2 is always 24 bits |
318 | // (due to the implicit bit), while _Exponent2 indicates a shift of at most 104 bits. |
319 | // 24 + 104 equals 128 equals 4 * 32, so we need exactly 4 32-bit elements. |
320 | // We use a little-endian representation, visualized like this: |
321 | |
322 | // << left shift << |
323 | // most significant |
324 | // _Data[3] _Data[2] _Data[1] _Data[0] |
325 | // least significant |
326 | // >> right shift >> |
327 | |
328 | constexpr uint32_t _Data_size = 4; |
329 | uint32_t _Data[_Data_size]{}; |
330 | |
331 | // _Maxidx is the index of the most significant nonzero element. |
332 | uint32_t _Maxidx = ((24 + static_cast<uint32_t>(_Exponent2) + 31) / 32) - 1; |
333 | _LIBCPP_ASSERT_INTERNAL(_Maxidx < _Data_size, "" ); |
334 | |
335 | const uint32_t _Bit_shift = static_cast<uint32_t>(_Exponent2) % 32; |
336 | if (_Bit_shift <= 8) { // _Mantissa2's 24 bits don't cross an element boundary |
337 | _Data[_Maxidx] = _Mantissa2 << _Bit_shift; |
338 | } else { // _Mantissa2's 24 bits cross an element boundary |
339 | _Data[_Maxidx - 1] = _Mantissa2 << _Bit_shift; |
340 | _Data[_Maxidx] = _Mantissa2 >> (32 - _Bit_shift); |
341 | } |
342 | |
343 | // If Ryu hasn't determined the total output length, we need to buffer the digits generated from right to left |
344 | // by long division. The largest possible float is: 340'282346638'528859811'704183484'516925440 |
345 | uint32_t _Blocks[4]; |
346 | int32_t _Filled_blocks = 0; |
347 | // From left to right, we're going to print: |
348 | // _Data[0] will be [1, 10] digits. |
349 | // Then if _Filled_blocks > 0: |
350 | // _Blocks[_Filled_blocks - 1], ..., _Blocks[0] will be 0-filled 9-digit blocks. |
351 | |
352 | if (_Maxidx != 0) { // If the integer is actually large, perform long division. |
353 | // Otherwise, skip to printing _Data[0]. |
354 | for (;;) { |
355 | // Loop invariant: _Maxidx != 0 (i.e. the integer is actually large) |
356 | |
357 | const uint32_t _Most_significant_elem = _Data[_Maxidx]; |
358 | const uint32_t _Initial_remainder = _Most_significant_elem % 1000000000; |
359 | const uint32_t _Initial_quotient = _Most_significant_elem / 1000000000; |
360 | _Data[_Maxidx] = _Initial_quotient; |
361 | uint64_t _Remainder = _Initial_remainder; |
362 | |
363 | // Process less significant elements. |
364 | uint32_t _Idx = _Maxidx; |
365 | do { |
366 | --_Idx; // Initially, _Remainder is at most 10^9 - 1. |
367 | |
368 | // Now, _Remainder is at most (10^9 - 1) * 2^32 + 2^32 - 1, simplified to 10^9 * 2^32 - 1. |
369 | _Remainder = (_Remainder << 32) | _Data[_Idx]; |
370 | |
371 | // floor((10^9 * 2^32 - 1) / 10^9) == 2^32 - 1, so uint32_t _Quotient is lossless. |
372 | const uint32_t _Quotient = static_cast<uint32_t>(__div1e9(_Remainder)); |
373 | |
374 | // _Remainder is at most 10^9 - 1 again. |
375 | // For uint32_t truncation, see the __mod1e9() comment in d2s_intrinsics.h. |
376 | _Remainder = static_cast<uint32_t>(_Remainder) - 1000000000u * _Quotient; |
377 | |
378 | _Data[_Idx] = _Quotient; |
379 | } while (_Idx != 0); |
380 | |
381 | // Store a 0-filled 9-digit block. |
382 | _Blocks[_Filled_blocks++] = static_cast<uint32_t>(_Remainder); |
383 | |
384 | if (_Initial_quotient == 0) { // Is the large integer shrinking? |
385 | --_Maxidx; // log2(10^9) is 29.9, so we can't shrink by more than one element. |
386 | if (_Maxidx == 0) { |
387 | break; // We've finished long division. Now we need to print _Data[0]. |
388 | } |
389 | } |
390 | } |
391 | } |
392 | |
393 | _LIBCPP_ASSERT_INTERNAL(_Data[0] != 0, "" ); |
394 | for (uint32_t _Idx = 1; _Idx < _Data_size; ++_Idx) { |
395 | _LIBCPP_ASSERT_INTERNAL(_Data[_Idx] == 0, "" ); |
396 | } |
397 | |
398 | const uint32_t _Data_olength = _Data[0] >= 1000000000 ? 10 : __decimalLength9(_Data[0]); |
399 | const uint32_t _Total_fixed_length = _Data_olength + 9 * _Filled_blocks; |
400 | |
401 | if (_Last - _First < static_cast<ptrdiff_t>(_Total_fixed_length)) { |
402 | return { _Last, errc::value_too_large }; |
403 | } |
404 | |
405 | char* _Result = _First; |
406 | |
407 | // Print _Data[0]. While it's up to 10 digits, |
408 | // which is more than Ryu generates, the code below can handle this. |
409 | __append_n_digits(_Data_olength, _Data[0], _Result); |
410 | _Result += _Data_olength; |
411 | |
412 | // Print 0-filled 9-digit blocks. |
413 | for (int32_t _Idx = _Filled_blocks - 1; _Idx >= 0; --_Idx) { |
414 | __append_nine_digits(_Blocks[_Idx], _Result); |
415 | _Result += 9; |
416 | } |
417 | |
418 | return { _Result, errc{} }; |
419 | } |
420 | |
421 | [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline to_chars_result __to_chars(char* const _First, char* const _Last, const __floating_decimal_32 __v, |
422 | chars_format _Fmt, const uint32_t __ieeeMantissa, const uint32_t __ieeeExponent) { |
423 | // Step 5: Print the decimal representation. |
424 | uint32_t _Output = __v.__mantissa; |
425 | int32_t _Ryu_exponent = __v.__exponent; |
426 | const uint32_t __olength = __decimalLength9(_Output); |
427 | int32_t _Scientific_exponent = _Ryu_exponent + static_cast<int32_t>(__olength) - 1; |
428 | |
429 | if (_Fmt == chars_format{}) { |
430 | int32_t _Lower; |
431 | int32_t _Upper; |
432 | |
433 | if (__olength == 1) { |
434 | // Value | Fixed | Scientific |
435 | // 1e-3 | "0.001" | "1e-03" |
436 | // 1e4 | "10000" | "1e+04" |
437 | _Lower = -3; |
438 | _Upper = 4; |
439 | } else { |
440 | // Value | Fixed | Scientific |
441 | // 1234e-7 | "0.0001234" | "1.234e-04" |
442 | // 1234e5 | "123400000" | "1.234e+08" |
443 | _Lower = -static_cast<int32_t>(__olength + 3); |
444 | _Upper = 5; |
445 | } |
446 | |
447 | if (_Lower <= _Ryu_exponent && _Ryu_exponent <= _Upper) { |
448 | _Fmt = chars_format::fixed; |
449 | } else { |
450 | _Fmt = chars_format::scientific; |
451 | } |
452 | } else if (_Fmt == chars_format::general) { |
453 | // C11 7.21.6.1 "The fprintf function"/8: |
454 | // "Let P equal [...] 6 if the precision is omitted [...]. |
455 | // Then, if a conversion with style E would have an exponent of X: |
456 | // - if P > X >= -4, the conversion is with style f [...]. |
457 | // - otherwise, the conversion is with style e [...]." |
458 | if (-4 <= _Scientific_exponent && _Scientific_exponent < 6) { |
459 | _Fmt = chars_format::fixed; |
460 | } else { |
461 | _Fmt = chars_format::scientific; |
462 | } |
463 | } |
464 | |
465 | if (_Fmt == chars_format::fixed) { |
466 | // Example: _Output == 1729, __olength == 4 |
467 | |
468 | // _Ryu_exponent | Printed | _Whole_digits | _Total_fixed_length | Notes |
469 | // --------------|----------|---------------|----------------------|--------------------------------------- |
470 | // 2 | 172900 | 6 | _Whole_digits | Ryu can't be used for printing |
471 | // 1 | 17290 | 5 | (sometimes adjusted) | when the trimmed digits are nonzero. |
472 | // --------------|----------|---------------|----------------------|--------------------------------------- |
473 | // 0 | 1729 | 4 | _Whole_digits | Unified length cases. |
474 | // --------------|----------|---------------|----------------------|--------------------------------------- |
475 | // -1 | 172.9 | 3 | __olength + 1 | This case can't happen for |
476 | // -2 | 17.29 | 2 | | __olength == 1, but no additional |
477 | // -3 | 1.729 | 1 | | code is needed to avoid it. |
478 | // --------------|----------|---------------|----------------------|--------------------------------------- |
479 | // -4 | 0.1729 | 0 | 2 - _Ryu_exponent | C11 7.21.6.1 "The fprintf function"/8: |
480 | // -5 | 0.01729 | -1 | | "If a decimal-point character appears, |
481 | // -6 | 0.001729 | -2 | | at least one digit appears before it." |
482 | |
483 | const int32_t _Whole_digits = static_cast<int32_t>(__olength) + _Ryu_exponent; |
484 | |
485 | uint32_t _Total_fixed_length; |
486 | if (_Ryu_exponent >= 0) { // cases "172900" and "1729" |
487 | _Total_fixed_length = static_cast<uint32_t>(_Whole_digits); |
488 | if (_Output == 1) { |
489 | // Rounding can affect the number of digits. |
490 | // For example, 1e11f is exactly "99999997952" which is 11 digits instead of 12. |
491 | // We can use a lookup table to detect this and adjust the total length. |
492 | static constexpr uint8_t _Adjustment[39] = { |
493 | 0,0,0,0,0,0,0,0,0,0,0,1,1,1,0,1,0,1,1,1,0,0,1,1,0,1,0,1,1,0,0,1,0,1,1,0,1,1,1 }; |
494 | _Total_fixed_length -= _Adjustment[_Ryu_exponent]; |
495 | // _Whole_digits doesn't need to be adjusted because these cases won't refer to it later. |
496 | } |
497 | } else if (_Whole_digits > 0) { // case "17.29" |
498 | _Total_fixed_length = __olength + 1; |
499 | } else { // case "0.001729" |
500 | _Total_fixed_length = static_cast<uint32_t>(2 - _Ryu_exponent); |
501 | } |
502 | |
503 | if (_Last - _First < static_cast<ptrdiff_t>(_Total_fixed_length)) { |
504 | return { _Last, errc::value_too_large }; |
505 | } |
506 | |
507 | char* _Mid; |
508 | if (_Ryu_exponent > 0) { // case "172900" |
509 | bool _Can_use_ryu; |
510 | |
511 | if (_Ryu_exponent > 10) { // 10^10 is the largest power of 10 that's exactly representable as a float. |
512 | _Can_use_ryu = false; |
513 | } else { |
514 | // Ryu generated X: __v.__mantissa * 10^_Ryu_exponent |
515 | // __v.__mantissa == 2^_Trailing_zero_bits * (__v.__mantissa >> _Trailing_zero_bits) |
516 | // 10^_Ryu_exponent == 2^_Ryu_exponent * 5^_Ryu_exponent |
517 | |
518 | // _Trailing_zero_bits is [0, 29] (aside: because 2^29 is the largest power of 2 |
519 | // with 9 decimal digits, which is float's round-trip limit.) |
520 | // _Ryu_exponent is [1, 10]. |
521 | // Normalization adds [2, 23] (aside: at least 2 because the pre-normalized mantissa is at least 5). |
522 | // This adds up to [3, 62], which is well below float's maximum binary exponent 127. |
523 | |
524 | // Therefore, we just need to consider (__v.__mantissa >> _Trailing_zero_bits) * 5^_Ryu_exponent. |
525 | |
526 | // If that product would exceed 24 bits, then X can't be exactly represented as a float. |
527 | // (That's not a problem for round-tripping, because X is close enough to the original float, |
528 | // but X isn't mathematically equal to the original float.) This requires a high-precision fallback. |
529 | |
530 | // If the product is 24 bits or smaller, then X can be exactly represented as a float (and we don't |
531 | // need to re-synthesize it; the original float must have been X, because Ryu wouldn't produce the |
532 | // same output for two different floats X and Y). This allows Ryu's output to be used (zero-filled). |
533 | |
534 | // (2^24 - 1) / 5^0 (for indexing), (2^24 - 1) / 5^1, ..., (2^24 - 1) / 5^10 |
535 | static constexpr uint32_t _Max_shifted_mantissa[11] = { |
536 | 16777215, 3355443, 671088, 134217, 26843, 5368, 1073, 214, 42, 8, 1 }; |
537 | |
538 | unsigned long _Trailing_zero_bits; |
539 | (void) _BitScanForward(&_Trailing_zero_bits, __v.__mantissa); // __v.__mantissa is guaranteed nonzero |
540 | const uint32_t _Shifted_mantissa = __v.__mantissa >> _Trailing_zero_bits; |
541 | _Can_use_ryu = _Shifted_mantissa <= _Max_shifted_mantissa[_Ryu_exponent]; |
542 | } |
543 | |
544 | if (!_Can_use_ryu) { |
545 | const uint32_t _Mantissa2 = __ieeeMantissa | (1u << __FLOAT_MANTISSA_BITS); // restore implicit bit |
546 | const int32_t _Exponent2 = static_cast<int32_t>(__ieeeExponent) |
547 | - __FLOAT_BIAS - __FLOAT_MANTISSA_BITS; // bias and normalization |
548 | |
549 | // Performance note: We've already called Ryu, so this will redundantly perform buffering and bounds checking. |
550 | return _Large_integer_to_chars(_First, _Last, _Mantissa2, _Exponent2); |
551 | } |
552 | |
553 | // _Can_use_ryu |
554 | // Print the decimal digits, left-aligned within [_First, _First + _Total_fixed_length). |
555 | _Mid = _First + __olength; |
556 | } else { // cases "1729", "17.29", and "0.001729" |
557 | // Print the decimal digits, right-aligned within [_First, _First + _Total_fixed_length). |
558 | _Mid = _First + _Total_fixed_length; |
559 | } |
560 | |
561 | while (_Output >= 10000) { |
562 | #ifdef __clang__ // TRANSITION, LLVM-38217 |
563 | const uint32_t __c = _Output - 10000 * (_Output / 10000); |
564 | #else |
565 | const uint32_t __c = _Output % 10000; |
566 | #endif |
567 | _Output /= 10000; |
568 | const uint32_t __c0 = (__c % 100) << 1; |
569 | const uint32_t __c1 = (__c / 100) << 1; |
570 | std::memcpy(_Mid -= 2, __DIGIT_TABLE + __c0, 2); |
571 | std::memcpy(_Mid -= 2, __DIGIT_TABLE + __c1, 2); |
572 | } |
573 | if (_Output >= 100) { |
574 | const uint32_t __c = (_Output % 100) << 1; |
575 | _Output /= 100; |
576 | std::memcpy(_Mid -= 2, __DIGIT_TABLE + __c, 2); |
577 | } |
578 | if (_Output >= 10) { |
579 | const uint32_t __c = _Output << 1; |
580 | std::memcpy(_Mid -= 2, __DIGIT_TABLE + __c, 2); |
581 | } else { |
582 | *--_Mid = static_cast<char>('0' + _Output); |
583 | } |
584 | |
585 | if (_Ryu_exponent > 0) { // case "172900" with _Can_use_ryu |
586 | // Performance note: it might be more efficient to do this immediately after setting _Mid. |
587 | std::memset(_First + __olength, '0', static_cast<size_t>(_Ryu_exponent)); |
588 | } else if (_Ryu_exponent == 0) { // case "1729" |
589 | // Done! |
590 | } else if (_Whole_digits > 0) { // case "17.29" |
591 | // Performance note: moving digits might not be optimal. |
592 | std::memmove(_First, _First + 1, static_cast<size_t>(_Whole_digits)); |
593 | _First[_Whole_digits] = '.'; |
594 | } else { // case "0.001729" |
595 | // Performance note: a larger memset() followed by overwriting '.' might be more efficient. |
596 | _First[0] = '0'; |
597 | _First[1] = '.'; |
598 | std::memset(_First + 2, '0', static_cast<size_t>(-_Whole_digits)); |
599 | } |
600 | |
601 | return { _First + _Total_fixed_length, errc{} }; |
602 | } |
603 | |
604 | const uint32_t _Total_scientific_length = |
605 | __olength + (__olength > 1) + 4; // digits + possible decimal point + scientific exponent |
606 | if (_Last - _First < static_cast<ptrdiff_t>(_Total_scientific_length)) { |
607 | return { _Last, errc::value_too_large }; |
608 | } |
609 | char* const __result = _First; |
610 | |
611 | // Print the decimal digits. |
612 | uint32_t __i = 0; |
613 | while (_Output >= 10000) { |
614 | #ifdef __clang__ // TRANSITION, LLVM-38217 |
615 | const uint32_t __c = _Output - 10000 * (_Output / 10000); |
616 | #else |
617 | const uint32_t __c = _Output % 10000; |
618 | #endif |
619 | _Output /= 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 (_Output >= 100) { |
627 | const uint32_t __c = (_Output % 100) << 1; |
628 | _Output /= 100; |
629 | std::memcpy(__result + __olength - __i - 1, __DIGIT_TABLE + __c, 2); |
630 | __i += 2; |
631 | } |
632 | if (_Output >= 10) { |
633 | const uint32_t __c = _Output << 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' + _Output); |
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 | std::memcpy(__result + __index, __DIGIT_TABLE + 2 * _Scientific_exponent, 2); |
660 | __index += 2; |
661 | |
662 | return { _First + _Total_scientific_length, errc{} }; |
663 | } |
664 | |
665 | [[nodiscard]] to_chars_result __f2s_buffered_n(char* const _First, char* const _Last, const float __f, |
666 | const chars_format _Fmt) { |
667 | |
668 | // Step 1: Decode the floating-point number, and unify normalized and subnormal cases. |
669 | const uint32_t __bits = __float_to_bits(__f); |
670 | |
671 | // Case distinction; exit early for the easy cases. |
672 | if (__bits == 0) { |
673 | if (_Fmt == chars_format::scientific) { |
674 | if (_Last - _First < 5) { |
675 | return { _Last, errc::value_too_large }; |
676 | } |
677 | |
678 | std::memcpy(_First, "0e+00" , 5); |
679 | |
680 | return { _First + 5, errc{} }; |
681 | } |
682 | |
683 | // Print "0" for chars_format::fixed, chars_format::general, and chars_format{}. |
684 | if (_First == _Last) { |
685 | return { _Last, errc::value_too_large }; |
686 | } |
687 | |
688 | *_First = '0'; |
689 | |
690 | return { _First + 1, errc{} }; |
691 | } |
692 | |
693 | // Decode __bits into mantissa and exponent. |
694 | const uint32_t __ieeeMantissa = __bits & ((1u << __FLOAT_MANTISSA_BITS) - 1); |
695 | const uint32_t __ieeeExponent = __bits >> __FLOAT_MANTISSA_BITS; |
696 | |
697 | // When _Fmt == chars_format::fixed and the floating-point number is a large integer, |
698 | // it's faster to skip Ryu and immediately print the integer exactly. |
699 | if (_Fmt == chars_format::fixed) { |
700 | const uint32_t _Mantissa2 = __ieeeMantissa | (1u << __FLOAT_MANTISSA_BITS); // restore implicit bit |
701 | const int32_t _Exponent2 = static_cast<int32_t>(__ieeeExponent) |
702 | - __FLOAT_BIAS - __FLOAT_MANTISSA_BITS; // bias and normalization |
703 | |
704 | // Normal values are equal to _Mantissa2 * 2^_Exponent2. |
705 | // (Subnormals are different, but they'll be rejected by the _Exponent2 test here, so they can be ignored.) |
706 | |
707 | if (_Exponent2 > 0) { |
708 | return _Large_integer_to_chars(_First, _Last, _Mantissa2, _Exponent2); |
709 | } |
710 | } |
711 | |
712 | const __floating_decimal_32 __v = __f2d(__ieeeMantissa, __ieeeExponent); |
713 | return __to_chars(_First, _Last, __v, _Fmt, __ieeeMantissa, __ieeeExponent); |
714 | } |
715 | |
716 | _LIBCPP_END_NAMESPACE_STD |
717 | |
718 | // clang-format on |
719 | |