xref: /freebsd/contrib/llvm-project/libcxx/src/ryu/f2s.cpp (revision 8ddb146abcdf061be9f2c0db7e391697dafad85c)
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 "__config"
43 #include "charconv"
44 
45 #include "include/ryu/common.h"
46 #include "include/ryu/d2fixed.h"
47 #include "include/ryu/d2s_intrinsics.h"
48 #include "include/ryu/digit_table.h"
49 #include "include/ryu/f2s.h"
50 #include "include/ryu/ryu.h"
51 
52 _LIBCPP_BEGIN_NAMESPACE_STD
53 
54 inline constexpr int __FLOAT_MANTISSA_BITS = 23;
55 inline constexpr int __FLOAT_EXPONENT_BITS = 8;
56 inline constexpr int __FLOAT_BIAS = 127;
57 
58 inline constexpr int __FLOAT_POW5_INV_BITCOUNT = 59;
59 inline constexpr uint64_t __FLOAT_POW5_INV_SPLIT[31] = {
60   576460752303423489u, 461168601842738791u, 368934881474191033u, 295147905179352826u,
61   472236648286964522u, 377789318629571618u, 302231454903657294u, 483570327845851670u,
62   386856262276681336u, 309485009821345069u, 495176015714152110u, 396140812571321688u,
63   316912650057057351u, 507060240091291761u, 405648192073033409u, 324518553658426727u,
64   519229685853482763u, 415383748682786211u, 332306998946228969u, 531691198313966350u,
65   425352958651173080u, 340282366920938464u, 544451787073501542u, 435561429658801234u,
66   348449143727040987u, 557518629963265579u, 446014903970612463u, 356811923176489971u,
67   570899077082383953u, 456719261665907162u, 365375409332725730u
68 };
69 inline constexpr int __FLOAT_POW5_BITCOUNT = 61;
70 inline constexpr uint64_t __FLOAT_POW5_SPLIT[47] = {
71   1152921504606846976u, 1441151880758558720u, 1801439850948198400u, 2251799813685248000u,
72   1407374883553280000u, 1759218604441600000u, 2199023255552000000u, 1374389534720000000u,
73   1717986918400000000u, 2147483648000000000u, 1342177280000000000u, 1677721600000000000u,
74   2097152000000000000u, 1310720000000000000u, 1638400000000000000u, 2048000000000000000u,
75   1280000000000000000u, 1600000000000000000u, 2000000000000000000u, 1250000000000000000u,
76   1562500000000000000u, 1953125000000000000u, 1220703125000000000u, 1525878906250000000u,
77   1907348632812500000u, 1192092895507812500u, 1490116119384765625u, 1862645149230957031u,
78   1164153218269348144u, 1455191522836685180u, 1818989403545856475u, 2273736754432320594u,
79   1421085471520200371u, 1776356839400250464u, 2220446049250313080u, 1387778780781445675u,
80   1734723475976807094u, 2168404344971008868u, 1355252715606880542u, 1694065894508600678u,
81   2117582368135750847u, 1323488980084844279u, 1654361225106055349u, 2067951531382569187u,
82   1292469707114105741u, 1615587133892632177u, 2019483917365790221u
83 };
84 
85 [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline uint32_t __pow5Factor(uint32_t __value) {
86   uint32_t __count = 0;
87   for (;;) {
88     _LIBCPP_ASSERT(__value != 0, "");
89     const uint32_t __q = __value / 5;
90     const uint32_t __r = __value % 5;
91     if (__r != 0) {
92       break;
93     }
94     __value = __q;
95     ++__count;
96   }
97   return __count;
98 }
99 
100 // Returns true if __value is divisible by 5^__p.
101 [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline bool __multipleOfPowerOf5(const uint32_t __value, const uint32_t __p) {
102   return __pow5Factor(__value) >= __p;
103 }
104 
105 // Returns true if __value is divisible by 2^__p.
106 [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline bool __multipleOfPowerOf2(const uint32_t __value, const uint32_t __p) {
107   _LIBCPP_ASSERT(__value != 0, "");
108   _LIBCPP_ASSERT(__p < 32, "");
109   // __builtin_ctz doesn't appear to be faster here.
110   return (__value & ((1u << __p) - 1)) == 0;
111 }
112 
113 [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline uint32_t __mulShift(const uint32_t __m, const uint64_t __factor, const int32_t __shift) {
114   _LIBCPP_ASSERT(__shift > 32, "");
115 
116   // The casts here help MSVC to avoid calls to the __allmul library
117   // function.
118   const uint32_t __factorLo = static_cast<uint32_t>(__factor);
119   const uint32_t __factorHi = static_cast<uint32_t>(__factor >> 32);
120   const uint64_t __bits0 = static_cast<uint64_t>(__m) * __factorLo;
121   const uint64_t __bits1 = static_cast<uint64_t>(__m) * __factorHi;
122 
123 #ifndef _LIBCPP_64_BIT
124   // On 32-bit platforms we can avoid a 64-bit shift-right since we only
125   // need the upper 32 bits of the result and the shift value is > 32.
126   const uint32_t __bits0Hi = static_cast<uint32_t>(__bits0 >> 32);
127   uint32_t __bits1Lo = static_cast<uint32_t>(__bits1);
128   uint32_t __bits1Hi = static_cast<uint32_t>(__bits1 >> 32);
129   __bits1Lo += __bits0Hi;
130   __bits1Hi += (__bits1Lo < __bits0Hi);
131   const int32_t __s = __shift - 32;
132   return (__bits1Hi << (32 - __s)) | (__bits1Lo >> __s);
133 #else // ^^^ 32-bit ^^^ / vvv 64-bit vvv
134   const uint64_t __sum = (__bits0 >> 32) + __bits1;
135   const uint64_t __shiftedSum = __sum >> (__shift - 32);
136   _LIBCPP_ASSERT(__shiftedSum <= UINT32_MAX, "");
137   return static_cast<uint32_t>(__shiftedSum);
138 #endif // ^^^ 64-bit ^^^
139 }
140 
141 [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline uint32_t __mulPow5InvDivPow2(const uint32_t __m, const uint32_t __q, const int32_t __j) {
142   return __mulShift(__m, __FLOAT_POW5_INV_SPLIT[__q], __j);
143 }
144 
145 [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline uint32_t __mulPow5divPow2(const uint32_t __m, const uint32_t __i, const int32_t __j) {
146   return __mulShift(__m, __FLOAT_POW5_SPLIT[__i], __j);
147 }
148 
149 // A floating decimal representing m * 10^e.
150 struct __floating_decimal_32 {
151   uint32_t __mantissa;
152   int32_t __exponent;
153 };
154 
155 [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline __floating_decimal_32 __f2d(const uint32_t __ieeeMantissa, const uint32_t __ieeeExponent) {
156   int32_t __e2;
157   uint32_t __m2;
158   if (__ieeeExponent == 0) {
159     // We subtract 2 so that the bounds computation has 2 additional bits.
160     __e2 = 1 - __FLOAT_BIAS - __FLOAT_MANTISSA_BITS - 2;
161     __m2 = __ieeeMantissa;
162   } else {
163     __e2 = static_cast<int32_t>(__ieeeExponent) - __FLOAT_BIAS - __FLOAT_MANTISSA_BITS - 2;
164     __m2 = (1u << __FLOAT_MANTISSA_BITS) | __ieeeMantissa;
165   }
166   const bool __even = (__m2 & 1) == 0;
167   const bool __acceptBounds = __even;
168 
169   // Step 2: Determine the interval of valid decimal representations.
170   const uint32_t __mv = 4 * __m2;
171   const uint32_t __mp = 4 * __m2 + 2;
172   // Implicit bool -> int conversion. True is 1, false is 0.
173   const uint32_t __mmShift = __ieeeMantissa != 0 || __ieeeExponent <= 1;
174   const uint32_t __mm = 4 * __m2 - 1 - __mmShift;
175 
176   // Step 3: Convert to a decimal power base using 64-bit arithmetic.
177   uint32_t __vr, __vp, __vm;
178   int32_t __e10;
179   bool __vmIsTrailingZeros = false;
180   bool __vrIsTrailingZeros = false;
181   uint8_t __lastRemovedDigit = 0;
182   if (__e2 >= 0) {
183     const uint32_t __q = __log10Pow2(__e2);
184     __e10 = static_cast<int32_t>(__q);
185     const int32_t __k = __FLOAT_POW5_INV_BITCOUNT + __pow5bits(static_cast<int32_t>(__q)) - 1;
186     const int32_t __i = -__e2 + static_cast<int32_t>(__q) + __k;
187     __vr = __mulPow5InvDivPow2(__mv, __q, __i);
188     __vp = __mulPow5InvDivPow2(__mp, __q, __i);
189     __vm = __mulPow5InvDivPow2(__mm, __q, __i);
190     if (__q != 0 && (__vp - 1) / 10 <= __vm / 10) {
191       // We need to know one removed digit even if we are not going to loop below. We could use
192       // __q = X - 1 above, except that would require 33 bits for the result, and we've found that
193       // 32-bit arithmetic is faster even on 64-bit machines.
194       const int32_t __l = __FLOAT_POW5_INV_BITCOUNT + __pow5bits(static_cast<int32_t>(__q - 1)) - 1;
195       __lastRemovedDigit = static_cast<uint8_t>(__mulPow5InvDivPow2(__mv, __q - 1,
196         -__e2 + static_cast<int32_t>(__q) - 1 + __l) % 10);
197     }
198     if (__q <= 9) {
199       // The largest power of 5 that fits in 24 bits is 5^10, but __q <= 9 seems to be safe as well.
200       // Only one of __mp, __mv, and __mm can be a multiple of 5, if any.
201       if (__mv % 5 == 0) {
202         __vrIsTrailingZeros = __multipleOfPowerOf5(__mv, __q);
203       } else if (__acceptBounds) {
204         __vmIsTrailingZeros = __multipleOfPowerOf5(__mm, __q);
205       } else {
206         __vp -= __multipleOfPowerOf5(__mp, __q);
207       }
208     }
209   } else {
210     const uint32_t __q = __log10Pow5(-__e2);
211     __e10 = static_cast<int32_t>(__q) + __e2;
212     const int32_t __i = -__e2 - static_cast<int32_t>(__q);
213     const int32_t __k = __pow5bits(__i) - __FLOAT_POW5_BITCOUNT;
214     int32_t __j = static_cast<int32_t>(__q) - __k;
215     __vr = __mulPow5divPow2(__mv, static_cast<uint32_t>(__i), __j);
216     __vp = __mulPow5divPow2(__mp, static_cast<uint32_t>(__i), __j);
217     __vm = __mulPow5divPow2(__mm, static_cast<uint32_t>(__i), __j);
218     if (__q != 0 && (__vp - 1) / 10 <= __vm / 10) {
219       __j = static_cast<int32_t>(__q) - 1 - (__pow5bits(__i + 1) - __FLOAT_POW5_BITCOUNT);
220       __lastRemovedDigit = static_cast<uint8_t>(__mulPow5divPow2(__mv, static_cast<uint32_t>(__i + 1), __j) % 10);
221     }
222     if (__q <= 1) {
223       // {__vr,__vp,__vm} is trailing zeros if {__mv,__mp,__mm} has at least __q trailing 0 bits.
224       // __mv = 4 * __m2, so it always has at least two trailing 0 bits.
225       __vrIsTrailingZeros = true;
226       if (__acceptBounds) {
227         // __mm = __mv - 1 - __mmShift, so it has 1 trailing 0 bit iff __mmShift == 1.
228         __vmIsTrailingZeros = __mmShift == 1;
229       } else {
230         // __mp = __mv + 2, so it always has at least one trailing 0 bit.
231         --__vp;
232       }
233     } else if (__q < 31) { // TRANSITION(ulfjack): Use a tighter bound here.
234       __vrIsTrailingZeros = __multipleOfPowerOf2(__mv, __q - 1);
235     }
236   }
237 
238   // Step 4: Find the shortest decimal representation in the interval of valid representations.
239   int32_t __removed = 0;
240   uint32_t _Output;
241   if (__vmIsTrailingZeros || __vrIsTrailingZeros) {
242     // General case, which happens rarely (~4.0%).
243     while (__vp / 10 > __vm / 10) {
244 #ifdef __clang__ // TRANSITION, LLVM-23106
245       __vmIsTrailingZeros &= __vm - (__vm / 10) * 10 == 0;
246 #else
247       __vmIsTrailingZeros &= __vm % 10 == 0;
248 #endif
249       __vrIsTrailingZeros &= __lastRemovedDigit == 0;
250       __lastRemovedDigit = static_cast<uint8_t>(__vr % 10);
251       __vr /= 10;
252       __vp /= 10;
253       __vm /= 10;
254       ++__removed;
255     }
256     if (__vmIsTrailingZeros) {
257       while (__vm % 10 == 0) {
258         __vrIsTrailingZeros &= __lastRemovedDigit == 0;
259         __lastRemovedDigit = static_cast<uint8_t>(__vr % 10);
260         __vr /= 10;
261         __vp /= 10;
262         __vm /= 10;
263         ++__removed;
264       }
265     }
266     if (__vrIsTrailingZeros && __lastRemovedDigit == 5 && __vr % 2 == 0) {
267       // Round even if the exact number is .....50..0.
268       __lastRemovedDigit = 4;
269     }
270     // We need to take __vr + 1 if __vr is outside bounds or we need to round up.
271     _Output = __vr + ((__vr == __vm && (!__acceptBounds || !__vmIsTrailingZeros)) || __lastRemovedDigit >= 5);
272   } else {
273     // Specialized for the common case (~96.0%). Percentages below are relative to this.
274     // Loop iterations below (approximately):
275     // 0: 13.6%, 1: 70.7%, 2: 14.1%, 3: 1.39%, 4: 0.14%, 5+: 0.01%
276     while (__vp / 10 > __vm / 10) {
277       __lastRemovedDigit = static_cast<uint8_t>(__vr % 10);
278       __vr /= 10;
279       __vp /= 10;
280       __vm /= 10;
281       ++__removed;
282     }
283     // We need to take __vr + 1 if __vr is outside bounds or we need to round up.
284     _Output = __vr + (__vr == __vm || __lastRemovedDigit >= 5);
285   }
286   const int32_t __exp = __e10 + __removed;
287 
288   __floating_decimal_32 __fd;
289   __fd.__exponent = __exp;
290   __fd.__mantissa = _Output;
291   return __fd;
292 }
293 
294 [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline to_chars_result _Large_integer_to_chars(char* const _First, char* const _Last,
295   const uint32_t _Mantissa2, const int32_t _Exponent2) {
296 
297   // Print the integer _Mantissa2 * 2^_Exponent2 exactly.
298 
299   // For nonzero integers, _Exponent2 >= -23. (The minimum value occurs when _Mantissa2 * 2^_Exponent2 is 1.
300   // In that case, _Mantissa2 is the implicit 1 bit followed by 23 zeros, so _Exponent2 is -23 to shift away
301   // the zeros.) The dense range of exactly representable integers has negative or zero exponents
302   // (as positive exponents make the range non-dense). For that dense range, Ryu will always be used:
303   // every digit is necessary to uniquely identify the value, so Ryu must print them all.
304 
305   // Positive exponents are the non-dense range of exactly representable integers.
306   // This contains all of the values for which Ryu can't be used (and a few Ryu-friendly values).
307 
308   // Performance note: Long division appears to be faster than losslessly widening float to double and calling
309   // __d2fixed_buffered_n(). If __f2fixed_buffered_n() is implemented, it might be faster than long division.
310 
311   _LIBCPP_ASSERT(_Exponent2 > 0, "");
312   _LIBCPP_ASSERT(_Exponent2 <= 104, ""); // because __ieeeExponent <= 254
313 
314   // Manually represent _Mantissa2 * 2^_Exponent2 as a large integer. _Mantissa2 is always 24 bits
315   // (due to the implicit bit), while _Exponent2 indicates a shift of at most 104 bits.
316   // 24 + 104 equals 128 equals 4 * 32, so we need exactly 4 32-bit elements.
317   // We use a little-endian representation, visualized like this:
318 
319   // << left shift <<
320   // most significant
321   // _Data[3] _Data[2] _Data[1] _Data[0]
322   //                   least significant
323   //                   >> right shift >>
324 
325   constexpr uint32_t _Data_size = 4;
326   uint32_t _Data[_Data_size]{};
327 
328   // _Maxidx is the index of the most significant nonzero element.
329   uint32_t _Maxidx = ((24 + static_cast<uint32_t>(_Exponent2) + 31) / 32) - 1;
330   _LIBCPP_ASSERT(_Maxidx < _Data_size, "");
331 
332   const uint32_t _Bit_shift = static_cast<uint32_t>(_Exponent2) % 32;
333   if (_Bit_shift <= 8) { // _Mantissa2's 24 bits don't cross an element boundary
334     _Data[_Maxidx] = _Mantissa2 << _Bit_shift;
335   } else { // _Mantissa2's 24 bits cross an element boundary
336     _Data[_Maxidx - 1] = _Mantissa2 << _Bit_shift;
337     _Data[_Maxidx] = _Mantissa2 >> (32 - _Bit_shift);
338   }
339 
340   // If Ryu hasn't determined the total output length, we need to buffer the digits generated from right to left
341   // by long division. The largest possible float is: 340'282346638'528859811'704183484'516925440
342   uint32_t _Blocks[4];
343   int32_t _Filled_blocks = 0;
344   // From left to right, we're going to print:
345   // _Data[0] will be [1, 10] digits.
346   // Then if _Filled_blocks > 0:
347   // _Blocks[_Filled_blocks - 1], ..., _Blocks[0] will be 0-filled 9-digit blocks.
348 
349   if (_Maxidx != 0) { // If the integer is actually large, perform long division.
350                       // Otherwise, skip to printing _Data[0].
351     for (;;) {
352       // Loop invariant: _Maxidx != 0 (i.e. the integer is actually large)
353 
354       const uint32_t _Most_significant_elem = _Data[_Maxidx];
355       const uint32_t _Initial_remainder = _Most_significant_elem % 1000000000;
356       const uint32_t _Initial_quotient = _Most_significant_elem / 1000000000;
357       _Data[_Maxidx] = _Initial_quotient;
358       uint64_t _Remainder = _Initial_remainder;
359 
360       // Process less significant elements.
361       uint32_t _Idx = _Maxidx;
362       do {
363         --_Idx; // Initially, _Remainder is at most 10^9 - 1.
364 
365         // Now, _Remainder is at most (10^9 - 1) * 2^32 + 2^32 - 1, simplified to 10^9 * 2^32 - 1.
366         _Remainder = (_Remainder << 32) | _Data[_Idx];
367 
368         // floor((10^9 * 2^32 - 1) / 10^9) == 2^32 - 1, so uint32_t _Quotient is lossless.
369         const uint32_t _Quotient = static_cast<uint32_t>(__div1e9(_Remainder));
370 
371         // _Remainder is at most 10^9 - 1 again.
372         // For uint32_t truncation, see the __mod1e9() comment in d2s_intrinsics.h.
373         _Remainder = static_cast<uint32_t>(_Remainder) - 1000000000u * _Quotient;
374 
375         _Data[_Idx] = _Quotient;
376       } while (_Idx != 0);
377 
378       // Store a 0-filled 9-digit block.
379       _Blocks[_Filled_blocks++] = static_cast<uint32_t>(_Remainder);
380 
381       if (_Initial_quotient == 0) { // Is the large integer shrinking?
382         --_Maxidx; // log2(10^9) is 29.9, so we can't shrink by more than one element.
383         if (_Maxidx == 0) {
384           break; // We've finished long division. Now we need to print _Data[0].
385         }
386       }
387     }
388   }
389 
390   _LIBCPP_ASSERT(_Data[0] != 0, "");
391   for (uint32_t _Idx = 1; _Idx < _Data_size; ++_Idx) {
392     _LIBCPP_ASSERT(_Data[_Idx] == 0, "");
393   }
394 
395   const uint32_t _Data_olength = _Data[0] >= 1000000000 ? 10 : __decimalLength9(_Data[0]);
396   const uint32_t _Total_fixed_length = _Data_olength + 9 * _Filled_blocks;
397 
398   if (_Last - _First < static_cast<ptrdiff_t>(_Total_fixed_length)) {
399     return { _Last, errc::value_too_large };
400   }
401 
402   char* _Result = _First;
403 
404   // Print _Data[0]. While it's up to 10 digits,
405   // which is more than Ryu generates, the code below can handle this.
406   __append_n_digits(_Data_olength, _Data[0], _Result);
407   _Result += _Data_olength;
408 
409   // Print 0-filled 9-digit blocks.
410   for (int32_t _Idx = _Filled_blocks - 1; _Idx >= 0; --_Idx) {
411     __append_nine_digits(_Blocks[_Idx], _Result);
412     _Result += 9;
413   }
414 
415   return { _Result, errc{} };
416 }
417 
418 [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline to_chars_result __to_chars(char* const _First, char* const _Last, const __floating_decimal_32 __v,
419   chars_format _Fmt, const uint32_t __ieeeMantissa, const uint32_t __ieeeExponent) {
420   // Step 5: Print the decimal representation.
421   uint32_t _Output = __v.__mantissa;
422   int32_t _Ryu_exponent = __v.__exponent;
423   const uint32_t __olength = __decimalLength9(_Output);
424   int32_t _Scientific_exponent = _Ryu_exponent + static_cast<int32_t>(__olength) - 1;
425 
426   if (_Fmt == chars_format{}) {
427     int32_t _Lower;
428     int32_t _Upper;
429 
430     if (__olength == 1) {
431       // Value | Fixed   | Scientific
432       // 1e-3  | "0.001" | "1e-03"
433       // 1e4   | "10000" | "1e+04"
434       _Lower = -3;
435       _Upper = 4;
436     } else {
437       // Value   | Fixed       | Scientific
438       // 1234e-7 | "0.0001234" | "1.234e-04"
439       // 1234e5  | "123400000" | "1.234e+08"
440       _Lower = -static_cast<int32_t>(__olength + 3);
441       _Upper = 5;
442     }
443 
444     if (_Lower <= _Ryu_exponent && _Ryu_exponent <= _Upper) {
445       _Fmt = chars_format::fixed;
446     } else {
447       _Fmt = chars_format::scientific;
448     }
449   } else if (_Fmt == chars_format::general) {
450     // C11 7.21.6.1 "The fprintf function"/8:
451     // "Let P equal [...] 6 if the precision is omitted [...].
452     // Then, if a conversion with style E would have an exponent of X:
453     // - if P > X >= -4, the conversion is with style f [...].
454     // - otherwise, the conversion is with style e [...]."
455     if (-4 <= _Scientific_exponent && _Scientific_exponent < 6) {
456       _Fmt = chars_format::fixed;
457     } else {
458       _Fmt = chars_format::scientific;
459     }
460   }
461 
462   if (_Fmt == chars_format::fixed) {
463     // Example: _Output == 1729, __olength == 4
464 
465     // _Ryu_exponent | Printed  | _Whole_digits | _Total_fixed_length  | Notes
466     // --------------|----------|---------------|----------------------|---------------------------------------
467     //             2 | 172900   |  6            | _Whole_digits        | Ryu can't be used for printing
468     //             1 | 17290    |  5            | (sometimes adjusted) | when the trimmed digits are nonzero.
469     // --------------|----------|---------------|----------------------|---------------------------------------
470     //             0 | 1729     |  4            | _Whole_digits        | Unified length cases.
471     // --------------|----------|---------------|----------------------|---------------------------------------
472     //            -1 | 172.9    |  3            | __olength + 1        | This case can't happen for
473     //            -2 | 17.29    |  2            |                      | __olength == 1, but no additional
474     //            -3 | 1.729    |  1            |                      | code is needed to avoid it.
475     // --------------|----------|---------------|----------------------|---------------------------------------
476     //            -4 | 0.1729   |  0            | 2 - _Ryu_exponent    | C11 7.21.6.1 "The fprintf function"/8:
477     //            -5 | 0.01729  | -1            |                      | "If a decimal-point character appears,
478     //            -6 | 0.001729 | -2            |                      | at least one digit appears before it."
479 
480     const int32_t _Whole_digits = static_cast<int32_t>(__olength) + _Ryu_exponent;
481 
482     uint32_t _Total_fixed_length;
483     if (_Ryu_exponent >= 0) { // cases "172900" and "1729"
484       _Total_fixed_length = static_cast<uint32_t>(_Whole_digits);
485       if (_Output == 1) {
486         // Rounding can affect the number of digits.
487         // For example, 1e11f is exactly "99999997952" which is 11 digits instead of 12.
488         // We can use a lookup table to detect this and adjust the total length.
489         static constexpr uint8_t _Adjustment[39] = {
490           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 };
491         _Total_fixed_length -= _Adjustment[_Ryu_exponent];
492         // _Whole_digits doesn't need to be adjusted because these cases won't refer to it later.
493       }
494     } else if (_Whole_digits > 0) { // case "17.29"
495       _Total_fixed_length = __olength + 1;
496     } else { // case "0.001729"
497       _Total_fixed_length = static_cast<uint32_t>(2 - _Ryu_exponent);
498     }
499 
500     if (_Last - _First < static_cast<ptrdiff_t>(_Total_fixed_length)) {
501       return { _Last, errc::value_too_large };
502     }
503 
504     char* _Mid;
505     if (_Ryu_exponent > 0) { // case "172900"
506       bool _Can_use_ryu;
507 
508       if (_Ryu_exponent > 10) { // 10^10 is the largest power of 10 that's exactly representable as a float.
509         _Can_use_ryu = false;
510       } else {
511         // Ryu generated X: __v.__mantissa * 10^_Ryu_exponent
512         // __v.__mantissa == 2^_Trailing_zero_bits * (__v.__mantissa >> _Trailing_zero_bits)
513         // 10^_Ryu_exponent == 2^_Ryu_exponent * 5^_Ryu_exponent
514 
515         // _Trailing_zero_bits is [0, 29] (aside: because 2^29 is the largest power of 2
516         // with 9 decimal digits, which is float's round-trip limit.)
517         // _Ryu_exponent is [1, 10].
518         // Normalization adds [2, 23] (aside: at least 2 because the pre-normalized mantissa is at least 5).
519         // This adds up to [3, 62], which is well below float's maximum binary exponent 127.
520 
521         // Therefore, we just need to consider (__v.__mantissa >> _Trailing_zero_bits) * 5^_Ryu_exponent.
522 
523         // If that product would exceed 24 bits, then X can't be exactly represented as a float.
524         // (That's not a problem for round-tripping, because X is close enough to the original float,
525         // but X isn't mathematically equal to the original float.) This requires a high-precision fallback.
526 
527         // If the product is 24 bits or smaller, then X can be exactly represented as a float (and we don't
528         // need to re-synthesize it; the original float must have been X, because Ryu wouldn't produce the
529         // same output for two different floats X and Y). This allows Ryu's output to be used (zero-filled).
530 
531         // (2^24 - 1) / 5^0 (for indexing), (2^24 - 1) / 5^1, ..., (2^24 - 1) / 5^10
532         static constexpr uint32_t _Max_shifted_mantissa[11] = {
533           16777215, 3355443, 671088, 134217, 26843, 5368, 1073, 214, 42, 8, 1 };
534 
535         unsigned long _Trailing_zero_bits;
536         (void) _BitScanForward(&_Trailing_zero_bits, __v.__mantissa); // __v.__mantissa is guaranteed nonzero
537         const uint32_t _Shifted_mantissa = __v.__mantissa >> _Trailing_zero_bits;
538         _Can_use_ryu = _Shifted_mantissa <= _Max_shifted_mantissa[_Ryu_exponent];
539       }
540 
541       if (!_Can_use_ryu) {
542         const uint32_t _Mantissa2 = __ieeeMantissa | (1u << __FLOAT_MANTISSA_BITS); // restore implicit bit
543         const int32_t _Exponent2 = static_cast<int32_t>(__ieeeExponent)
544           - __FLOAT_BIAS - __FLOAT_MANTISSA_BITS; // bias and normalization
545 
546         // Performance note: We've already called Ryu, so this will redundantly perform buffering and bounds checking.
547         return _Large_integer_to_chars(_First, _Last, _Mantissa2, _Exponent2);
548       }
549 
550       // _Can_use_ryu
551       // Print the decimal digits, left-aligned within [_First, _First + _Total_fixed_length).
552       _Mid = _First + __olength;
553     } else { // cases "1729", "17.29", and "0.001729"
554       // Print the decimal digits, right-aligned within [_First, _First + _Total_fixed_length).
555       _Mid = _First + _Total_fixed_length;
556     }
557 
558     while (_Output >= 10000) {
559 #ifdef __clang__ // TRANSITION, LLVM-38217
560       const uint32_t __c = _Output - 10000 * (_Output / 10000);
561 #else
562       const uint32_t __c = _Output % 10000;
563 #endif
564       _Output /= 10000;
565       const uint32_t __c0 = (__c % 100) << 1;
566       const uint32_t __c1 = (__c / 100) << 1;
567       _VSTD::memcpy(_Mid -= 2, __DIGIT_TABLE + __c0, 2);
568       _VSTD::memcpy(_Mid -= 2, __DIGIT_TABLE + __c1, 2);
569     }
570     if (_Output >= 100) {
571       const uint32_t __c = (_Output % 100) << 1;
572       _Output /= 100;
573       _VSTD::memcpy(_Mid -= 2, __DIGIT_TABLE + __c, 2);
574     }
575     if (_Output >= 10) {
576       const uint32_t __c = _Output << 1;
577       _VSTD::memcpy(_Mid -= 2, __DIGIT_TABLE + __c, 2);
578     } else {
579       *--_Mid = static_cast<char>('0' + _Output);
580     }
581 
582     if (_Ryu_exponent > 0) { // case "172900" with _Can_use_ryu
583       // Performance note: it might be more efficient to do this immediately after setting _Mid.
584       _VSTD::memset(_First + __olength, '0', static_cast<size_t>(_Ryu_exponent));
585     } else if (_Ryu_exponent == 0) { // case "1729"
586       // Done!
587     } else if (_Whole_digits > 0) { // case "17.29"
588       // Performance note: moving digits might not be optimal.
589       _VSTD::memmove(_First, _First + 1, static_cast<size_t>(_Whole_digits));
590       _First[_Whole_digits] = '.';
591     } else { // case "0.001729"
592       // Performance note: a larger memset() followed by overwriting '.' might be more efficient.
593       _First[0] = '0';
594       _First[1] = '.';
595       _VSTD::memset(_First + 2, '0', static_cast<size_t>(-_Whole_digits));
596     }
597 
598     return { _First + _Total_fixed_length, errc{} };
599   }
600 
601   const uint32_t _Total_scientific_length =
602     __olength + (__olength > 1) + 4; // digits + possible decimal point + scientific exponent
603   if (_Last - _First < static_cast<ptrdiff_t>(_Total_scientific_length)) {
604     return { _Last, errc::value_too_large };
605   }
606   char* const __result = _First;
607 
608   // Print the decimal digits.
609   uint32_t __i = 0;
610   while (_Output >= 10000) {
611 #ifdef __clang__ // TRANSITION, LLVM-38217
612     const uint32_t __c = _Output - 10000 * (_Output / 10000);
613 #else
614     const uint32_t __c = _Output % 10000;
615 #endif
616     _Output /= 10000;
617     const uint32_t __c0 = (__c % 100) << 1;
618     const uint32_t __c1 = (__c / 100) << 1;
619     _VSTD::memcpy(__result + __olength - __i - 1, __DIGIT_TABLE + __c0, 2);
620     _VSTD::memcpy(__result + __olength - __i - 3, __DIGIT_TABLE + __c1, 2);
621     __i += 4;
622   }
623   if (_Output >= 100) {
624     const uint32_t __c = (_Output % 100) << 1;
625     _Output /= 100;
626     _VSTD::memcpy(__result + __olength - __i - 1, __DIGIT_TABLE + __c, 2);
627     __i += 2;
628   }
629   if (_Output >= 10) {
630     const uint32_t __c = _Output << 1;
631     // We can't use memcpy here: the decimal dot goes between these two digits.
632     __result[2] = __DIGIT_TABLE[__c + 1];
633     __result[0] = __DIGIT_TABLE[__c];
634   } else {
635     __result[0] = static_cast<char>('0' + _Output);
636   }
637 
638   // Print decimal point if needed.
639   uint32_t __index;
640   if (__olength > 1) {
641     __result[1] = '.';
642     __index = __olength + 1;
643   } else {
644     __index = 1;
645   }
646 
647   // Print the exponent.
648   __result[__index++] = 'e';
649   if (_Scientific_exponent < 0) {
650     __result[__index++] = '-';
651     _Scientific_exponent = -_Scientific_exponent;
652   } else {
653     __result[__index++] = '+';
654   }
655 
656   _VSTD::memcpy(__result + __index, __DIGIT_TABLE + 2 * _Scientific_exponent, 2);
657   __index += 2;
658 
659   return { _First + _Total_scientific_length, errc{} };
660 }
661 
662 [[nodiscard]] to_chars_result __f2s_buffered_n(char* const _First, char* const _Last, const float __f,
663   const chars_format _Fmt) {
664 
665   // Step 1: Decode the floating-point number, and unify normalized and subnormal cases.
666   const uint32_t __bits = __float_to_bits(__f);
667 
668   // Case distinction; exit early for the easy cases.
669   if (__bits == 0) {
670     if (_Fmt == chars_format::scientific) {
671       if (_Last - _First < 5) {
672         return { _Last, errc::value_too_large };
673       }
674 
675       _VSTD::memcpy(_First, "0e+00", 5);
676 
677       return { _First + 5, errc{} };
678     }
679 
680     // Print "0" for chars_format::fixed, chars_format::general, and chars_format{}.
681     if (_First == _Last) {
682       return { _Last, errc::value_too_large };
683     }
684 
685     *_First = '0';
686 
687     return { _First + 1, errc{} };
688   }
689 
690   // Decode __bits into mantissa and exponent.
691   const uint32_t __ieeeMantissa = __bits & ((1u << __FLOAT_MANTISSA_BITS) - 1);
692   const uint32_t __ieeeExponent = __bits >> __FLOAT_MANTISSA_BITS;
693 
694   // When _Fmt == chars_format::fixed and the floating-point number is a large integer,
695   // it's faster to skip Ryu and immediately print the integer exactly.
696   if (_Fmt == chars_format::fixed) {
697     const uint32_t _Mantissa2 = __ieeeMantissa | (1u << __FLOAT_MANTISSA_BITS); // restore implicit bit
698     const int32_t _Exponent2 = static_cast<int32_t>(__ieeeExponent)
699       - __FLOAT_BIAS - __FLOAT_MANTISSA_BITS; // bias and normalization
700 
701     // Normal values are equal to _Mantissa2 * 2^_Exponent2.
702     // (Subnormals are different, but they'll be rejected by the _Exponent2 test here, so they can be ignored.)
703 
704     if (_Exponent2 > 0) {
705       return _Large_integer_to_chars(_First, _Last, _Mantissa2, _Exponent2);
706     }
707   }
708 
709   const __floating_decimal_32 __v = __f2d(__ieeeMantissa, __ieeeExponent);
710   return __to_chars(_First, _Last, __v, _Fmt, __ieeeMantissa, __ieeeExponent);
711 }
712 
713 _LIBCPP_END_NAMESPACE_STD
714 
715 // clang-format on
716