xref: /freebsd/contrib/llvm-project/libcxx/src/ryu/d2s.cpp (revision 5f757f3ff9144b609b3c433dfd370cc6bdc191ad)
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 
46 #include "include/ryu/common.h"
47 #include "include/ryu/d2fixed.h"
48 #include "include/ryu/d2s.h"
49 #include "include/ryu/d2s_full_table.h"
50 #include "include/ryu/d2s_intrinsics.h"
51 #include "include/ryu/digit_table.h"
52 #include "include/ryu/ryu.h"
53 
54 _LIBCPP_BEGIN_NAMESPACE_STD
55 
56 // We need a 64x128-bit multiplication and a subsequent 128-bit shift.
57 // Multiplication:
58 //   The 64-bit factor is variable and passed in, the 128-bit factor comes
59 //   from a lookup table. We know that the 64-bit factor only has 55
60 //   significant bits (i.e., the 9 topmost bits are zeros). The 128-bit
61 //   factor only has 124 significant bits (i.e., the 4 topmost bits are
62 //   zeros).
63 // Shift:
64 //   In principle, the multiplication result requires 55 + 124 = 179 bits to
65 //   represent. However, we then shift this value to the right by __j, which is
66 //   at least __j >= 115, so the result is guaranteed to fit into 179 - 115 = 64
67 //   bits. This means that we only need the topmost 64 significant bits of
68 //   the 64x128-bit multiplication.
69 //
70 // There are several ways to do this:
71 // 1. Best case: the compiler exposes a 128-bit type.
72 //    We perform two 64x64-bit multiplications, add the higher 64 bits of the
73 //    lower result to the higher result, and shift by __j - 64 bits.
74 //
75 //    We explicitly cast from 64-bit to 128-bit, so the compiler can tell
76 //    that these are only 64-bit inputs, and can map these to the best
77 //    possible sequence of assembly instructions.
78 //    x64 machines happen to have matching assembly instructions for
79 //    64x64-bit multiplications and 128-bit shifts.
80 //
81 // 2. Second best case: the compiler exposes intrinsics for the x64 assembly
82 //    instructions mentioned in 1.
83 //
84 // 3. We only have 64x64 bit instructions that return the lower 64 bits of
85 //    the result, i.e., we have to use plain C.
86 //    Our inputs are less than the full width, so we have three options:
87 //    a. Ignore this fact and just implement the intrinsics manually.
88 //    b. Split both into 31-bit pieces, which guarantees no internal overflow,
89 //       but requires extra work upfront (unless we change the lookup table).
90 //    c. Split only the first factor into 31-bit pieces, which also guarantees
91 //       no internal overflow, but requires extra work since the intermediate
92 //       results are not perfectly aligned.
93 #ifdef _LIBCPP_INTRINSIC128
94 
95 [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline uint64_t __mulShift(const uint64_t __m, const uint64_t* const __mul, const int32_t __j) {
96   // __m is maximum 55 bits
97   uint64_t __high1;                                               // 128
98   const uint64_t __low1 = __ryu_umul128(__m, __mul[1], &__high1); // 64
99   uint64_t __high0;                                               // 64
100   (void) __ryu_umul128(__m, __mul[0], &__high0);                  // 0
101   const uint64_t __sum = __high0 + __low1;
102   if (__sum < __high0) {
103     ++__high1; // overflow into __high1
104   }
105   return __ryu_shiftright128(__sum, __high1, static_cast<uint32_t>(__j - 64));
106 }
107 
108 [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline uint64_t __mulShiftAll(const uint64_t __m, const uint64_t* const __mul, const int32_t __j,
109   uint64_t* const __vp, uint64_t* const __vm, const uint32_t __mmShift) {
110   *__vp = __mulShift(4 * __m + 2, __mul, __j);
111   *__vm = __mulShift(4 * __m - 1 - __mmShift, __mul, __j);
112   return __mulShift(4 * __m, __mul, __j);
113 }
114 
115 #else // ^^^ intrinsics available ^^^ / vvv intrinsics unavailable vvv
116 
117 [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline _LIBCPP_ALWAYS_INLINE uint64_t __mulShiftAll(uint64_t __m, const uint64_t* const __mul, const int32_t __j,
118   uint64_t* const __vp, uint64_t* const __vm, const uint32_t __mmShift) { // TRANSITION, VSO-634761
119   __m <<= 1;
120   // __m is maximum 55 bits
121   uint64_t __tmp;
122   const uint64_t __lo = __ryu_umul128(__m, __mul[0], &__tmp);
123   uint64_t __hi;
124   const uint64_t __mid = __tmp + __ryu_umul128(__m, __mul[1], &__hi);
125   __hi += __mid < __tmp; // overflow into __hi
126 
127   const uint64_t __lo2 = __lo + __mul[0];
128   const uint64_t __mid2 = __mid + __mul[1] + (__lo2 < __lo);
129   const uint64_t __hi2 = __hi + (__mid2 < __mid);
130   *__vp = __ryu_shiftright128(__mid2, __hi2, static_cast<uint32_t>(__j - 64 - 1));
131 
132   if (__mmShift == 1) {
133     const uint64_t __lo3 = __lo - __mul[0];
134     const uint64_t __mid3 = __mid - __mul[1] - (__lo3 > __lo);
135     const uint64_t __hi3 = __hi - (__mid3 > __mid);
136     *__vm = __ryu_shiftright128(__mid3, __hi3, static_cast<uint32_t>(__j - 64 - 1));
137   } else {
138     const uint64_t __lo3 = __lo + __lo;
139     const uint64_t __mid3 = __mid + __mid + (__lo3 < __lo);
140     const uint64_t __hi3 = __hi + __hi + (__mid3 < __mid);
141     const uint64_t __lo4 = __lo3 - __mul[0];
142     const uint64_t __mid4 = __mid3 - __mul[1] - (__lo4 > __lo3);
143     const uint64_t __hi4 = __hi3 - (__mid4 > __mid3);
144     *__vm = __ryu_shiftright128(__mid4, __hi4, static_cast<uint32_t>(__j - 64));
145   }
146 
147   return __ryu_shiftright128(__mid, __hi, static_cast<uint32_t>(__j - 64 - 1));
148 }
149 
150 #endif // ^^^ intrinsics unavailable ^^^
151 
152 [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline uint32_t __decimalLength17(const uint64_t __v) {
153   // This is slightly faster than a loop.
154   // The average output length is 16.38 digits, so we check high-to-low.
155   // Function precondition: __v is not an 18, 19, or 20-digit number.
156   // (17 digits are sufficient for round-tripping.)
157   _LIBCPP_ASSERT_INTERNAL(__v < 100000000000000000u, "");
158   if (__v >= 10000000000000000u) { return 17; }
159   if (__v >= 1000000000000000u) { return 16; }
160   if (__v >= 100000000000000u) { return 15; }
161   if (__v >= 10000000000000u) { return 14; }
162   if (__v >= 1000000000000u) { return 13; }
163   if (__v >= 100000000000u) { return 12; }
164   if (__v >= 10000000000u) { return 11; }
165   if (__v >= 1000000000u) { return 10; }
166   if (__v >= 100000000u) { return 9; }
167   if (__v >= 10000000u) { return 8; }
168   if (__v >= 1000000u) { return 7; }
169   if (__v >= 100000u) { return 6; }
170   if (__v >= 10000u) { return 5; }
171   if (__v >= 1000u) { return 4; }
172   if (__v >= 100u) { return 3; }
173   if (__v >= 10u) { return 2; }
174   return 1;
175 }
176 
177 // A floating decimal representing m * 10^e.
178 struct __floating_decimal_64 {
179   uint64_t __mantissa;
180   int32_t __exponent;
181 };
182 
183 [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline __floating_decimal_64 __d2d(const uint64_t __ieeeMantissa, const uint32_t __ieeeExponent) {
184   int32_t __e2;
185   uint64_t __m2;
186   if (__ieeeExponent == 0) {
187     // We subtract 2 so that the bounds computation has 2 additional bits.
188     __e2 = 1 - __DOUBLE_BIAS - __DOUBLE_MANTISSA_BITS - 2;
189     __m2 = __ieeeMantissa;
190   } else {
191     __e2 = static_cast<int32_t>(__ieeeExponent) - __DOUBLE_BIAS - __DOUBLE_MANTISSA_BITS - 2;
192     __m2 = (1ull << __DOUBLE_MANTISSA_BITS) | __ieeeMantissa;
193   }
194   const bool __even = (__m2 & 1) == 0;
195   const bool __acceptBounds = __even;
196 
197   // Step 2: Determine the interval of valid decimal representations.
198   const uint64_t __mv = 4 * __m2;
199   // Implicit bool -> int conversion. True is 1, false is 0.
200   const uint32_t __mmShift = __ieeeMantissa != 0 || __ieeeExponent <= 1;
201   // We would compute __mp and __mm like this:
202   // uint64_t __mp = 4 * __m2 + 2;
203   // uint64_t __mm = __mv - 1 - __mmShift;
204 
205   // Step 3: Convert to a decimal power base using 128-bit arithmetic.
206   uint64_t __vr, __vp, __vm;
207   int32_t __e10;
208   bool __vmIsTrailingZeros = false;
209   bool __vrIsTrailingZeros = false;
210   if (__e2 >= 0) {
211     // I tried special-casing __q == 0, but there was no effect on performance.
212     // This expression is slightly faster than max(0, __log10Pow2(__e2) - 1).
213     const uint32_t __q = __log10Pow2(__e2) - (__e2 > 3);
214     __e10 = static_cast<int32_t>(__q);
215     const int32_t __k = __DOUBLE_POW5_INV_BITCOUNT + __pow5bits(static_cast<int32_t>(__q)) - 1;
216     const int32_t __i = -__e2 + static_cast<int32_t>(__q) + __k;
217     __vr = __mulShiftAll(__m2, __DOUBLE_POW5_INV_SPLIT[__q], __i, &__vp, &__vm, __mmShift);
218     if (__q <= 21) {
219       // This should use __q <= 22, but I think 21 is also safe. Smaller values
220       // may still be safe, but it's more difficult to reason about them.
221       // Only one of __mp, __mv, and __mm can be a multiple of 5, if any.
222       const uint32_t __mvMod5 = static_cast<uint32_t>(__mv) - 5 * static_cast<uint32_t>(__div5(__mv));
223       if (__mvMod5 == 0) {
224         __vrIsTrailingZeros = __multipleOfPowerOf5(__mv, __q);
225       } else if (__acceptBounds) {
226         // Same as min(__e2 + (~__mm & 1), __pow5Factor(__mm)) >= __q
227         // <=> __e2 + (~__mm & 1) >= __q && __pow5Factor(__mm) >= __q
228         // <=> true && __pow5Factor(__mm) >= __q, since __e2 >= __q.
229         __vmIsTrailingZeros = __multipleOfPowerOf5(__mv - 1 - __mmShift, __q);
230       } else {
231         // Same as min(__e2 + 1, __pow5Factor(__mp)) >= __q.
232         __vp -= __multipleOfPowerOf5(__mv + 2, __q);
233       }
234     }
235   } else {
236     // This expression is slightly faster than max(0, __log10Pow5(-__e2) - 1).
237     const uint32_t __q = __log10Pow5(-__e2) - (-__e2 > 1);
238     __e10 = static_cast<int32_t>(__q) + __e2;
239     const int32_t __i = -__e2 - static_cast<int32_t>(__q);
240     const int32_t __k = __pow5bits(__i) - __DOUBLE_POW5_BITCOUNT;
241     const int32_t __j = static_cast<int32_t>(__q) - __k;
242     __vr = __mulShiftAll(__m2, __DOUBLE_POW5_SPLIT[__i], __j, &__vp, &__vm, __mmShift);
243     if (__q <= 1) {
244       // {__vr,__vp,__vm} is trailing zeros if {__mv,__mp,__mm} has at least __q trailing 0 bits.
245       // __mv = 4 * __m2, so it always has at least two trailing 0 bits.
246       __vrIsTrailingZeros = true;
247       if (__acceptBounds) {
248         // __mm = __mv - 1 - __mmShift, so it has 1 trailing 0 bit iff __mmShift == 1.
249         __vmIsTrailingZeros = __mmShift == 1;
250       } else {
251         // __mp = __mv + 2, so it always has at least one trailing 0 bit.
252         --__vp;
253       }
254     } else if (__q < 63) { // TRANSITION(ulfjack): Use a tighter bound here.
255       // We need to compute min(ntz(__mv), __pow5Factor(__mv) - __e2) >= __q - 1
256       // <=> ntz(__mv) >= __q - 1 && __pow5Factor(__mv) - __e2 >= __q - 1
257       // <=> ntz(__mv) >= __q - 1 (__e2 is negative and -__e2 >= __q)
258       // <=> (__mv & ((1 << (__q - 1)) - 1)) == 0
259       // We also need to make sure that the left shift does not overflow.
260       __vrIsTrailingZeros = __multipleOfPowerOf2(__mv, __q - 1);
261     }
262   }
263 
264   // Step 4: Find the shortest decimal representation in the interval of valid representations.
265   int32_t __removed = 0;
266   uint8_t __lastRemovedDigit = 0;
267   uint64_t _Output;
268   // On average, we remove ~2 digits.
269   if (__vmIsTrailingZeros || __vrIsTrailingZeros) {
270     // General case, which happens rarely (~0.7%).
271     for (;;) {
272       const uint64_t __vpDiv10 = __div10(__vp);
273       const uint64_t __vmDiv10 = __div10(__vm);
274       if (__vpDiv10 <= __vmDiv10) {
275         break;
276       }
277       const uint32_t __vmMod10 = static_cast<uint32_t>(__vm) - 10 * static_cast<uint32_t>(__vmDiv10);
278       const uint64_t __vrDiv10 = __div10(__vr);
279       const uint32_t __vrMod10 = static_cast<uint32_t>(__vr) - 10 * static_cast<uint32_t>(__vrDiv10);
280       __vmIsTrailingZeros &= __vmMod10 == 0;
281       __vrIsTrailingZeros &= __lastRemovedDigit == 0;
282       __lastRemovedDigit = static_cast<uint8_t>(__vrMod10);
283       __vr = __vrDiv10;
284       __vp = __vpDiv10;
285       __vm = __vmDiv10;
286       ++__removed;
287     }
288     if (__vmIsTrailingZeros) {
289       for (;;) {
290         const uint64_t __vmDiv10 = __div10(__vm);
291         const uint32_t __vmMod10 = static_cast<uint32_t>(__vm) - 10 * static_cast<uint32_t>(__vmDiv10);
292         if (__vmMod10 != 0) {
293           break;
294         }
295         const uint64_t __vpDiv10 = __div10(__vp);
296         const uint64_t __vrDiv10 = __div10(__vr);
297         const uint32_t __vrMod10 = static_cast<uint32_t>(__vr) - 10 * static_cast<uint32_t>(__vrDiv10);
298         __vrIsTrailingZeros &= __lastRemovedDigit == 0;
299         __lastRemovedDigit = static_cast<uint8_t>(__vrMod10);
300         __vr = __vrDiv10;
301         __vp = __vpDiv10;
302         __vm = __vmDiv10;
303         ++__removed;
304       }
305     }
306     if (__vrIsTrailingZeros && __lastRemovedDigit == 5 && __vr % 2 == 0) {
307       // Round even if the exact number is .....50..0.
308       __lastRemovedDigit = 4;
309     }
310     // We need to take __vr + 1 if __vr is outside bounds or we need to round up.
311     _Output = __vr + ((__vr == __vm && (!__acceptBounds || !__vmIsTrailingZeros)) || __lastRemovedDigit >= 5);
312   } else {
313     // Specialized for the common case (~99.3%). Percentages below are relative to this.
314     bool __roundUp = false;
315     const uint64_t __vpDiv100 = __div100(__vp);
316     const uint64_t __vmDiv100 = __div100(__vm);
317     if (__vpDiv100 > __vmDiv100) { // Optimization: remove two digits at a time (~86.2%).
318       const uint64_t __vrDiv100 = __div100(__vr);
319       const uint32_t __vrMod100 = static_cast<uint32_t>(__vr) - 100 * static_cast<uint32_t>(__vrDiv100);
320       __roundUp = __vrMod100 >= 50;
321       __vr = __vrDiv100;
322       __vp = __vpDiv100;
323       __vm = __vmDiv100;
324       __removed += 2;
325     }
326     // Loop iterations below (approximately), without optimization above:
327     // 0: 0.03%, 1: 13.8%, 2: 70.6%, 3: 14.0%, 4: 1.40%, 5: 0.14%, 6+: 0.02%
328     // Loop iterations below (approximately), with optimization above:
329     // 0: 70.6%, 1: 27.8%, 2: 1.40%, 3: 0.14%, 4+: 0.02%
330     for (;;) {
331       const uint64_t __vpDiv10 = __div10(__vp);
332       const uint64_t __vmDiv10 = __div10(__vm);
333       if (__vpDiv10 <= __vmDiv10) {
334         break;
335       }
336       const uint64_t __vrDiv10 = __div10(__vr);
337       const uint32_t __vrMod10 = static_cast<uint32_t>(__vr) - 10 * static_cast<uint32_t>(__vrDiv10);
338       __roundUp = __vrMod10 >= 5;
339       __vr = __vrDiv10;
340       __vp = __vpDiv10;
341       __vm = __vmDiv10;
342       ++__removed;
343     }
344     // We need to take __vr + 1 if __vr is outside bounds or we need to round up.
345     _Output = __vr + (__vr == __vm || __roundUp);
346   }
347   const int32_t __exp = __e10 + __removed;
348 
349   __floating_decimal_64 __fd;
350   __fd.__exponent = __exp;
351   __fd.__mantissa = _Output;
352   return __fd;
353 }
354 
355 [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline to_chars_result __to_chars(char* const _First, char* const _Last, const __floating_decimal_64 __v,
356   chars_format _Fmt, const double __f) {
357   // Step 5: Print the decimal representation.
358   uint64_t _Output = __v.__mantissa;
359   int32_t _Ryu_exponent = __v.__exponent;
360   const uint32_t __olength = __decimalLength17(_Output);
361   int32_t _Scientific_exponent = _Ryu_exponent + static_cast<int32_t>(__olength) - 1;
362 
363   if (_Fmt == chars_format{}) {
364     int32_t _Lower;
365     int32_t _Upper;
366 
367     if (__olength == 1) {
368       // Value | Fixed   | Scientific
369       // 1e-3  | "0.001" | "1e-03"
370       // 1e4   | "10000" | "1e+04"
371       _Lower = -3;
372       _Upper = 4;
373     } else {
374       // Value   | Fixed       | Scientific
375       // 1234e-7 | "0.0001234" | "1.234e-04"
376       // 1234e5  | "123400000" | "1.234e+08"
377       _Lower = -static_cast<int32_t>(__olength + 3);
378       _Upper = 5;
379     }
380 
381     if (_Lower <= _Ryu_exponent && _Ryu_exponent <= _Upper) {
382       _Fmt = chars_format::fixed;
383     } else {
384       _Fmt = chars_format::scientific;
385     }
386   } else if (_Fmt == chars_format::general) {
387     // C11 7.21.6.1 "The fprintf function"/8:
388     // "Let P equal [...] 6 if the precision is omitted [...].
389     // Then, if a conversion with style E would have an exponent of X:
390     // - if P > X >= -4, the conversion is with style f [...].
391     // - otherwise, the conversion is with style e [...]."
392     if (-4 <= _Scientific_exponent && _Scientific_exponent < 6) {
393       _Fmt = chars_format::fixed;
394     } else {
395       _Fmt = chars_format::scientific;
396     }
397   }
398 
399   if (_Fmt == chars_format::fixed) {
400     // Example: _Output == 1729, __olength == 4
401 
402     // _Ryu_exponent | Printed  | _Whole_digits | _Total_fixed_length  | Notes
403     // --------------|----------|---------------|----------------------|---------------------------------------
404     //             2 | 172900   |  6            | _Whole_digits        | Ryu can't be used for printing
405     //             1 | 17290    |  5            | (sometimes adjusted) | when the trimmed digits are nonzero.
406     // --------------|----------|---------------|----------------------|---------------------------------------
407     //             0 | 1729     |  4            | _Whole_digits        | Unified length cases.
408     // --------------|----------|---------------|----------------------|---------------------------------------
409     //            -1 | 172.9    |  3            | __olength + 1        | This case can't happen for
410     //            -2 | 17.29    |  2            |                      | __olength == 1, but no additional
411     //            -3 | 1.729    |  1            |                      | code is needed to avoid it.
412     // --------------|----------|---------------|----------------------|---------------------------------------
413     //            -4 | 0.1729   |  0            | 2 - _Ryu_exponent    | C11 7.21.6.1 "The fprintf function"/8:
414     //            -5 | 0.01729  | -1            |                      | "If a decimal-point character appears,
415     //            -6 | 0.001729 | -2            |                      | at least one digit appears before it."
416 
417     const int32_t _Whole_digits = static_cast<int32_t>(__olength) + _Ryu_exponent;
418 
419     uint32_t _Total_fixed_length;
420     if (_Ryu_exponent >= 0) { // cases "172900" and "1729"
421       _Total_fixed_length = static_cast<uint32_t>(_Whole_digits);
422       if (_Output == 1) {
423         // Rounding can affect the number of digits.
424         // For example, 1e23 is exactly "99999999999999991611392" which is 23 digits instead of 24.
425         // We can use a lookup table to detect this and adjust the total length.
426         static constexpr uint8_t _Adjustment[309] = {
427           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,
428           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,
429           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,
430           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,
431           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,
432           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,
433           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 };
434         _Total_fixed_length -= _Adjustment[_Ryu_exponent];
435         // _Whole_digits doesn't need to be adjusted because these cases won't refer to it later.
436       }
437     } else if (_Whole_digits > 0) { // case "17.29"
438       _Total_fixed_length = __olength + 1;
439     } else { // case "0.001729"
440       _Total_fixed_length = static_cast<uint32_t>(2 - _Ryu_exponent);
441     }
442 
443     if (_Last - _First < static_cast<ptrdiff_t>(_Total_fixed_length)) {
444       return { _Last, errc::value_too_large };
445     }
446 
447     char* _Mid;
448     if (_Ryu_exponent > 0) { // case "172900"
449       bool _Can_use_ryu;
450 
451       if (_Ryu_exponent > 22) { // 10^22 is the largest power of 10 that's exactly representable as a double.
452         _Can_use_ryu = false;
453       } else {
454         // Ryu generated X: __v.__mantissa * 10^_Ryu_exponent
455         // __v.__mantissa == 2^_Trailing_zero_bits * (__v.__mantissa >> _Trailing_zero_bits)
456         // 10^_Ryu_exponent == 2^_Ryu_exponent * 5^_Ryu_exponent
457 
458         // _Trailing_zero_bits is [0, 56] (aside: because 2^56 is the largest power of 2
459         // with 17 decimal digits, which is double's round-trip limit.)
460         // _Ryu_exponent is [1, 22].
461         // Normalization adds [2, 52] (aside: at least 2 because the pre-normalized mantissa is at least 5).
462         // This adds up to [3, 130], which is well below double's maximum binary exponent 1023.
463 
464         // Therefore, we just need to consider (__v.__mantissa >> _Trailing_zero_bits) * 5^_Ryu_exponent.
465 
466         // If that product would exceed 53 bits, then X can't be exactly represented as a double.
467         // (That's not a problem for round-tripping, because X is close enough to the original double,
468         // but X isn't mathematically equal to the original double.) This requires a high-precision fallback.
469 
470         // If the product is 53 bits or smaller, then X can be exactly represented as a double (and we don't
471         // need to re-synthesize it; the original double must have been X, because Ryu wouldn't produce the
472         // same output for two different doubles X and Y). This allows Ryu's output to be used (zero-filled).
473 
474         // (2^53 - 1) / 5^0 (for indexing), (2^53 - 1) / 5^1, ..., (2^53 - 1) / 5^22
475         static constexpr uint64_t _Max_shifted_mantissa[23] = {
476           9007199254740991u, 1801439850948198u, 360287970189639u, 72057594037927u, 14411518807585u,
477           2882303761517u, 576460752303u, 115292150460u, 23058430092u, 4611686018u, 922337203u, 184467440u,
478           36893488u, 7378697u, 1475739u, 295147u, 59029u, 11805u, 2361u, 472u, 94u, 18u, 3u };
479 
480         unsigned long _Trailing_zero_bits;
481 #ifdef _LIBCPP_HAS_BITSCAN64
482         (void) _BitScanForward64(&_Trailing_zero_bits, __v.__mantissa); // __v.__mantissa is guaranteed nonzero
483 #else // ^^^ 64-bit ^^^ / vvv 32-bit vvv
484         const uint32_t _Low_mantissa = static_cast<uint32_t>(__v.__mantissa);
485         if (_Low_mantissa != 0) {
486           (void) _BitScanForward(&_Trailing_zero_bits, _Low_mantissa);
487         } else {
488           const uint32_t _High_mantissa = static_cast<uint32_t>(__v.__mantissa >> 32); // nonzero here
489           (void) _BitScanForward(&_Trailing_zero_bits, _High_mantissa);
490           _Trailing_zero_bits += 32;
491         }
492 #endif // ^^^ 32-bit ^^^
493         const uint64_t _Shifted_mantissa = __v.__mantissa >> _Trailing_zero_bits;
494         _Can_use_ryu = _Shifted_mantissa <= _Max_shifted_mantissa[_Ryu_exponent];
495       }
496 
497       if (!_Can_use_ryu) {
498         // Print the integer exactly.
499         // Performance note: This will redundantly perform bounds checking.
500         // Performance note: This will redundantly decompose the IEEE representation.
501         return __d2fixed_buffered_n(_First, _Last, __f, 0);
502       }
503 
504       // _Can_use_ryu
505       // Print the decimal digits, left-aligned within [_First, _First + _Total_fixed_length).
506       _Mid = _First + __olength;
507     } else { // cases "1729", "17.29", and "0.001729"
508       // Print the decimal digits, right-aligned within [_First, _First + _Total_fixed_length).
509       _Mid = _First + _Total_fixed_length;
510     }
511 
512     // We prefer 32-bit operations, even on 64-bit platforms.
513     // We have at most 17 digits, and uint32_t can store 9 digits.
514     // If _Output doesn't fit into uint32_t, we cut off 8 digits,
515     // so the rest will fit into uint32_t.
516     if ((_Output >> 32) != 0) {
517       // Expensive 64-bit division.
518       const uint64_t __q = __div1e8(_Output);
519       uint32_t __output2 = static_cast<uint32_t>(_Output - 100000000 * __q);
520       _Output = __q;
521 
522       const uint32_t __c = __output2 % 10000;
523       __output2 /= 10000;
524       const uint32_t __d = __output2 % 10000;
525       const uint32_t __c0 = (__c % 100) << 1;
526       const uint32_t __c1 = (__c / 100) << 1;
527       const uint32_t __d0 = (__d % 100) << 1;
528       const uint32_t __d1 = (__d / 100) << 1;
529 
530       std::memcpy(_Mid -= 2, __DIGIT_TABLE + __c0, 2);
531       std::memcpy(_Mid -= 2, __DIGIT_TABLE + __c1, 2);
532       std::memcpy(_Mid -= 2, __DIGIT_TABLE + __d0, 2);
533       std::memcpy(_Mid -= 2, __DIGIT_TABLE + __d1, 2);
534     }
535     uint32_t __output2 = static_cast<uint32_t>(_Output);
536     while (__output2 >= 10000) {
537 #ifdef __clang__ // TRANSITION, LLVM-38217
538       const uint32_t __c = __output2 - 10000 * (__output2 / 10000);
539 #else
540       const uint32_t __c = __output2 % 10000;
541 #endif
542       __output2 /= 10000;
543       const uint32_t __c0 = (__c % 100) << 1;
544       const uint32_t __c1 = (__c / 100) << 1;
545       std::memcpy(_Mid -= 2, __DIGIT_TABLE + __c0, 2);
546       std::memcpy(_Mid -= 2, __DIGIT_TABLE + __c1, 2);
547     }
548     if (__output2 >= 100) {
549       const uint32_t __c = (__output2 % 100) << 1;
550       __output2 /= 100;
551       std::memcpy(_Mid -= 2, __DIGIT_TABLE + __c, 2);
552     }
553     if (__output2 >= 10) {
554       const uint32_t __c = __output2 << 1;
555       std::memcpy(_Mid -= 2, __DIGIT_TABLE + __c, 2);
556     } else {
557       *--_Mid = static_cast<char>('0' + __output2);
558     }
559 
560     if (_Ryu_exponent > 0) { // case "172900" with _Can_use_ryu
561       // Performance note: it might be more efficient to do this immediately after setting _Mid.
562       std::memset(_First + __olength, '0', static_cast<size_t>(_Ryu_exponent));
563     } else if (_Ryu_exponent == 0) { // case "1729"
564       // Done!
565     } else if (_Whole_digits > 0) { // case "17.29"
566       // Performance note: moving digits might not be optimal.
567       std::memmove(_First, _First + 1, static_cast<size_t>(_Whole_digits));
568       _First[_Whole_digits] = '.';
569     } else { // case "0.001729"
570       // Performance note: a larger memset() followed by overwriting '.' might be more efficient.
571       _First[0] = '0';
572       _First[1] = '.';
573       std::memset(_First + 2, '0', static_cast<size_t>(-_Whole_digits));
574     }
575 
576     return { _First + _Total_fixed_length, errc{} };
577   }
578 
579   const uint32_t _Total_scientific_length = __olength + (__olength > 1) // digits + possible decimal point
580     + (-100 < _Scientific_exponent && _Scientific_exponent < 100 ? 4 : 5); // + scientific exponent
581   if (_Last - _First < static_cast<ptrdiff_t>(_Total_scientific_length)) {
582     return { _Last, errc::value_too_large };
583   }
584   char* const __result = _First;
585 
586   // Print the decimal digits.
587   uint32_t __i = 0;
588   // We prefer 32-bit operations, even on 64-bit platforms.
589   // We have at most 17 digits, and uint32_t can store 9 digits.
590   // If _Output doesn't fit into uint32_t, we cut off 8 digits,
591   // so the rest will fit into uint32_t.
592   if ((_Output >> 32) != 0) {
593     // Expensive 64-bit division.
594     const uint64_t __q = __div1e8(_Output);
595     uint32_t __output2 = static_cast<uint32_t>(_Output) - 100000000 * static_cast<uint32_t>(__q);
596     _Output = __q;
597 
598     const uint32_t __c = __output2 % 10000;
599     __output2 /= 10000;
600     const uint32_t __d = __output2 % 10000;
601     const uint32_t __c0 = (__c % 100) << 1;
602     const uint32_t __c1 = (__c / 100) << 1;
603     const uint32_t __d0 = (__d % 100) << 1;
604     const uint32_t __d1 = (__d / 100) << 1;
605     std::memcpy(__result + __olength - __i - 1, __DIGIT_TABLE + __c0, 2);
606     std::memcpy(__result + __olength - __i - 3, __DIGIT_TABLE + __c1, 2);
607     std::memcpy(__result + __olength - __i - 5, __DIGIT_TABLE + __d0, 2);
608     std::memcpy(__result + __olength - __i - 7, __DIGIT_TABLE + __d1, 2);
609     __i += 8;
610   }
611   uint32_t __output2 = static_cast<uint32_t>(_Output);
612   while (__output2 >= 10000) {
613 #ifdef __clang__ // TRANSITION, LLVM-38217
614     const uint32_t __c = __output2 - 10000 * (__output2 / 10000);
615 #else
616     const uint32_t __c = __output2 % 10000;
617 #endif
618     __output2 /= 10000;
619     const uint32_t __c0 = (__c % 100) << 1;
620     const uint32_t __c1 = (__c / 100) << 1;
621     std::memcpy(__result + __olength - __i - 1, __DIGIT_TABLE + __c0, 2);
622     std::memcpy(__result + __olength - __i - 3, __DIGIT_TABLE + __c1, 2);
623     __i += 4;
624   }
625   if (__output2 >= 100) {
626     const uint32_t __c = (__output2 % 100) << 1;
627     __output2 /= 100;
628     std::memcpy(__result + __olength - __i - 1, __DIGIT_TABLE + __c, 2);
629     __i += 2;
630   }
631   if (__output2 >= 10) {
632     const uint32_t __c = __output2 << 1;
633     // We can't use memcpy here: the decimal dot goes between these two digits.
634     __result[2] = __DIGIT_TABLE[__c + 1];
635     __result[0] = __DIGIT_TABLE[__c];
636   } else {
637     __result[0] = static_cast<char>('0' + __output2);
638   }
639 
640   // Print decimal point if needed.
641   uint32_t __index;
642   if (__olength > 1) {
643     __result[1] = '.';
644     __index = __olength + 1;
645   } else {
646     __index = 1;
647   }
648 
649   // Print the exponent.
650   __result[__index++] = 'e';
651   if (_Scientific_exponent < 0) {
652     __result[__index++] = '-';
653     _Scientific_exponent = -_Scientific_exponent;
654   } else {
655     __result[__index++] = '+';
656   }
657 
658   if (_Scientific_exponent >= 100) {
659     const int32_t __c = _Scientific_exponent % 10;
660     std::memcpy(__result + __index, __DIGIT_TABLE + 2 * (_Scientific_exponent / 10), 2);
661     __result[__index + 2] = static_cast<char>('0' + __c);
662     __index += 3;
663   } else {
664     std::memcpy(__result + __index, __DIGIT_TABLE + 2 * _Scientific_exponent, 2);
665     __index += 2;
666   }
667 
668   return { _First + _Total_scientific_length, errc{} };
669 }
670 
671 [[nodiscard]] _LIBCPP_HIDE_FROM_ABI inline bool __d2d_small_int(const uint64_t __ieeeMantissa, const uint32_t __ieeeExponent,
672   __floating_decimal_64* const __v) {
673   const uint64_t __m2 = (1ull << __DOUBLE_MANTISSA_BITS) | __ieeeMantissa;
674   const int32_t __e2 = static_cast<int32_t>(__ieeeExponent) - __DOUBLE_BIAS - __DOUBLE_MANTISSA_BITS;
675 
676   if (__e2 > 0) {
677     // f = __m2 * 2^__e2 >= 2^53 is an integer.
678     // Ignore this case for now.
679     return false;
680   }
681 
682   if (__e2 < -52) {
683     // f < 1.
684     return false;
685   }
686 
687   // Since 2^52 <= __m2 < 2^53 and 0 <= -__e2 <= 52: 1 <= f = __m2 / 2^-__e2 < 2^53.
688   // Test if the lower -__e2 bits of the significand are 0, i.e. whether the fraction is 0.
689   const uint64_t __mask = (1ull << -__e2) - 1;
690   const uint64_t __fraction = __m2 & __mask;
691   if (__fraction != 0) {
692     return false;
693   }
694 
695   // f is an integer in the range [1, 2^53).
696   // Note: __mantissa might contain trailing (decimal) 0's.
697   // Note: since 2^53 < 10^16, there is no need to adjust __decimalLength17().
698   __v->__mantissa = __m2 >> -__e2;
699   __v->__exponent = 0;
700   return true;
701 }
702 
703 [[nodiscard]] to_chars_result __d2s_buffered_n(char* const _First, char* const _Last, const double __f,
704   const chars_format _Fmt) {
705 
706   // Step 1: Decode the floating-point number, and unify normalized and subnormal cases.
707   const uint64_t __bits = __double_to_bits(__f);
708 
709   // Case distinction; exit early for the easy cases.
710   if (__bits == 0) {
711     if (_Fmt == chars_format::scientific) {
712       if (_Last - _First < 5) {
713         return { _Last, errc::value_too_large };
714       }
715 
716       std::memcpy(_First, "0e+00", 5);
717 
718       return { _First + 5, errc{} };
719     }
720 
721     // Print "0" for chars_format::fixed, chars_format::general, and chars_format{}.
722     if (_First == _Last) {
723       return { _Last, errc::value_too_large };
724     }
725 
726     *_First = '0';
727 
728     return { _First + 1, errc{} };
729   }
730 
731   // Decode __bits into mantissa and exponent.
732   const uint64_t __ieeeMantissa = __bits & ((1ull << __DOUBLE_MANTISSA_BITS) - 1);
733   const uint32_t __ieeeExponent = static_cast<uint32_t>(__bits >> __DOUBLE_MANTISSA_BITS);
734 
735   if (_Fmt == chars_format::fixed) {
736     // const uint64_t _Mantissa2 = __ieeeMantissa | (1ull << __DOUBLE_MANTISSA_BITS); // restore implicit bit
737     const int32_t _Exponent2 = static_cast<int32_t>(__ieeeExponent)
738       - __DOUBLE_BIAS - __DOUBLE_MANTISSA_BITS; // bias and normalization
739 
740     // Normal values are equal to _Mantissa2 * 2^_Exponent2.
741     // (Subnormals are different, but they'll be rejected by the _Exponent2 test here, so they can be ignored.)
742 
743     // For nonzero integers, _Exponent2 >= -52. (The minimum value occurs when _Mantissa2 * 2^_Exponent2 is 1.
744     // In that case, _Mantissa2 is the implicit 1 bit followed by 52 zeros, so _Exponent2 is -52 to shift away
745     // the zeros.) The dense range of exactly representable integers has negative or zero exponents
746     // (as positive exponents make the range non-dense). For that dense range, Ryu will always be used:
747     // every digit is necessary to uniquely identify the value, so Ryu must print them all.
748 
749     // Positive exponents are the non-dense range of exactly representable integers. This contains all of the values
750     // for which Ryu can't be used (and a few Ryu-friendly values). We can save time by detecting positive
751     // exponents here and skipping Ryu. Calling __d2fixed_buffered_n() with precision 0 is valid for all integers
752     // (so it's okay if we call it with a Ryu-friendly value).
753     if (_Exponent2 > 0) {
754       return __d2fixed_buffered_n(_First, _Last, __f, 0);
755     }
756   }
757 
758   __floating_decimal_64 __v;
759   const bool __isSmallInt = __d2d_small_int(__ieeeMantissa, __ieeeExponent, &__v);
760   if (__isSmallInt) {
761     // For small integers in the range [1, 2^53), __v.__mantissa might contain trailing (decimal) zeros.
762     // For scientific notation we need to move these zeros into the exponent.
763     // (This is not needed for fixed-point notation, so it might be beneficial to trim
764     // trailing zeros in __to_chars only if needed - once fixed-point notation output is implemented.)
765     for (;;) {
766       const uint64_t __q = __div10(__v.__mantissa);
767       const uint32_t __r = static_cast<uint32_t>(__v.__mantissa) - 10 * static_cast<uint32_t>(__q);
768       if (__r != 0) {
769         break;
770       }
771       __v.__mantissa = __q;
772       ++__v.__exponent;
773     }
774   } else {
775     __v = __d2d(__ieeeMantissa, __ieeeExponent);
776   }
777 
778   return __to_chars(_First, _Last, __v, _Fmt, __f);
779 }
780 
781 _LIBCPP_END_NAMESPACE_STD
782 
783 // clang-format on
784