/src/mysql-server/strings/dtoa.cc
Line | Count | Source |
1 | | /* Copyright (c) 2007, 2025, Oracle and/or its affiliates. |
2 | | |
3 | | This program is free software; you can redistribute it and/or modify |
4 | | it under the terms of the GNU General Public License, version 2.0, |
5 | | as published by the Free Software Foundation. |
6 | | |
7 | | This program is designed to work with certain software (including |
8 | | but not limited to OpenSSL) that is licensed under separate terms, |
9 | | as designated in a particular file or component or in included license |
10 | | documentation. The authors of MySQL hereby grant you an additional |
11 | | permission to link the program and your derivative works with the |
12 | | separately licensed software that they have either included with |
13 | | the program or referenced in the documentation. |
14 | | |
15 | | Without limiting anything contained in the foregoing, this file, |
16 | | which is part of C Driver for MySQL (Connector/C), is also subject to the |
17 | | Universal FOSS Exception, version 1.0, a copy of which can be found at |
18 | | http://oss.oracle.com/licenses/universal-foss-exception. |
19 | | |
20 | | This program is distributed in the hope that it will be useful, |
21 | | but WITHOUT ANY WARRANTY; without even the implied warranty of |
22 | | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
23 | | GNU General Public License, version 2.0, for more details. |
24 | | |
25 | | You should have received a copy of the GNU General Public License |
26 | | along with this program; if not, write to the Free Software |
27 | | Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA */ |
28 | | |
29 | | /**************************************************************** |
30 | | |
31 | | This file incorporates work covered by the following copyright and |
32 | | permission notice: |
33 | | |
34 | | The author of this software is David M. Gay. |
35 | | |
36 | | Copyright (c) 1991, 2000, 2001 by Lucent Technologies. |
37 | | |
38 | | Permission to use, copy, modify, and distribute this software for any |
39 | | purpose without fee is hereby granted, provided that this entire notice |
40 | | is included in all copies of any software which is or includes a copy |
41 | | or modification of this software and in all copies of the supporting |
42 | | documentation for such software. |
43 | | |
44 | | THIS SOFTWARE IS BEING PROVIDED "AS IS", WITHOUT ANY EXPRESS OR IMPLIED |
45 | | WARRANTY. IN PARTICULAR, NEITHER THE AUTHOR NOR LUCENT MAKES ANY |
46 | | REPRESENTATION OR WARRANTY OF ANY KIND CONCERNING THE MERCHANTABILITY |
47 | | OF THIS SOFTWARE OR ITS FITNESS FOR ANY PARTICULAR PURPOSE. |
48 | | |
49 | | ***************************************************************/ |
50 | | |
51 | | #include "my_config.h" |
52 | | |
53 | | #include "mysql/strings/dtoa.h" |
54 | | |
55 | | #include <algorithm> |
56 | | #include <cassert> |
57 | | #include <cerrno> |
58 | | #include <cfloat> |
59 | | #include <cstdint> |
60 | | #include <cstdlib> |
61 | | #include <cstring> |
62 | | #include <limits> |
63 | | |
64 | | #include "my_pointer_arithmetic.h" |
65 | | #include "template_utils.h" |
66 | | |
67 | | #ifndef EOVERFLOW |
68 | | #define EOVERFLOW 84 |
69 | | #endif |
70 | | |
71 | | /** |
72 | | Appears to suffice to not call malloc() in most cases. |
73 | | @todo |
74 | | see if it is possible to get rid of malloc(). |
75 | | this constant is sufficient to avoid malloc() on all inputs I have tried. |
76 | | */ |
77 | | #define DTOA_BUFF_SIZE (460 * sizeof(void *)) |
78 | | |
79 | | /* Magic value returned by dtoa() to indicate overflow */ |
80 | 0 | #define DTOA_OVERFLOW 9999 |
81 | | |
82 | | static double my_strtod_int(const char *, const char **, int *, char *, size_t); |
83 | | static char *dtoa(double, int, int, int *, int *, char **, char *, size_t); |
84 | | static void dtoa_free(char *, char *, size_t); |
85 | | |
86 | | /** |
87 | | @brief |
88 | | Converts a given floating point number to a zero-terminated string |
89 | | representation using the 'f' format. |
90 | | |
91 | | @details |
92 | | This function is a wrapper around dtoa() to do the same as |
93 | | sprintf(to, "%-.*f", precision, x), though the conversion is usually more |
94 | | precise. The only difference is in handling [-,+]infinity and nan values, |
95 | | in which case we print '0\0' to the output string and indicate an overflow. |
96 | | |
97 | | @param x the input floating point number. |
98 | | @param precision the number of digits after the decimal point. |
99 | | All properties of sprintf() apply: |
100 | | - if the number of significant digits after the decimal |
101 | | point is less than precision, the resulting string is |
102 | | right-padded with zeros |
103 | | - if the precision is 0, no decimal point appears |
104 | | - if a decimal point appears, at least one digit appears |
105 | | before it |
106 | | @param to pointer to the output buffer. The longest string which |
107 | | my_fcvt() can return is FLOATING_POINT_BUFFER bytes |
108 | | (including the terminating '\0'). |
109 | | @param error if not NULL, points to a location where the status of |
110 | | conversion is stored upon return. |
111 | | false successful conversion |
112 | | true the input number is [-,+]infinity or nan. |
113 | | The output string in this case is always '0'. |
114 | | @param shorten Whether to minimize the number of significant digits. If |
115 | | true, write only the minimum number of digits necessary to |
116 | | reproduce the double value when parsing the string. If |
117 | | false, zeros are added to the end to reach the precision |
118 | | limit. |
119 | | |
120 | | @return number of written characters (excluding terminating '\0') |
121 | | */ |
122 | | |
123 | | static size_t my_fcvt_internal(double x, int precision, bool shorten, char *to, |
124 | 0 | bool *error) { |
125 | 0 | int decpt, sign, len, i; |
126 | 0 | char *res, *src, *end, *dst = to; |
127 | 0 | char buf[DTOA_BUFF_SIZE]; |
128 | 0 | assert(precision >= 0 && precision < DECIMAL_NOT_SPECIFIED && to != nullptr); |
129 | |
|
130 | 0 | res = dtoa(x, 5, precision, &decpt, &sign, &end, buf, sizeof(buf)); |
131 | |
|
132 | 0 | if (decpt == DTOA_OVERFLOW) { |
133 | 0 | dtoa_free(res, buf, sizeof(buf)); |
134 | 0 | *to++ = '0'; |
135 | 0 | *to = '\0'; |
136 | 0 | if (error != nullptr) *error = true; |
137 | 0 | return 1; |
138 | 0 | } |
139 | | |
140 | 0 | src = res; |
141 | 0 | len = (int)(end - src); |
142 | |
|
143 | 0 | if (sign) *dst++ = '-'; |
144 | |
|
145 | 0 | if (decpt <= 0) { |
146 | 0 | *dst++ = '0'; |
147 | 0 | *dst++ = '.'; |
148 | 0 | for (i = decpt; i < 0; i++) *dst++ = '0'; |
149 | 0 | } |
150 | |
|
151 | 0 | for (i = 1; i <= len; i++) { |
152 | 0 | *dst++ = *src++; |
153 | 0 | if (i == decpt && i < len) *dst++ = '.'; |
154 | 0 | } |
155 | 0 | while (i++ <= decpt) *dst++ = '0'; |
156 | |
|
157 | 0 | if (precision > 0 && !shorten) { |
158 | 0 | if (len <= decpt) *dst++ = '.'; |
159 | |
|
160 | 0 | for (i = precision - std::max(0, (len - decpt)); i > 0; i--) *dst++ = '0'; |
161 | 0 | } |
162 | |
|
163 | 0 | *dst = '\0'; |
164 | 0 | if (error != nullptr) *error = false; |
165 | |
|
166 | 0 | dtoa_free(res, buf, sizeof(buf)); |
167 | |
|
168 | 0 | return dst - to; |
169 | 0 | } |
170 | | |
171 | | /** |
172 | | @brief |
173 | | Converts a given floating point number to a zero-terminated string |
174 | | representation using the 'f' format. |
175 | | |
176 | | @details |
177 | | This function is a wrapper around dtoa() to do the same as |
178 | | sprintf(to, "%-.*f", precision, x), though the conversion is usually more |
179 | | precise. The only difference is in handling [-,+]infinity and nan values, |
180 | | in which case we print '0\0' to the output string and indicate an overflow. |
181 | | |
182 | | @param x the input floating point number. |
183 | | @param precision the number of digits after the decimal point. |
184 | | All properties of sprintf() apply: |
185 | | - if the number of significant digits after the decimal |
186 | | point is less than precision, the resulting string is |
187 | | right-padded with zeros |
188 | | - if the precision is 0, no decimal point appears |
189 | | - if a decimal point appears, at least one digit appears |
190 | | before it |
191 | | @param to pointer to the output buffer. The longest string which |
192 | | my_fcvt() can return is FLOATING_POINT_BUFFER bytes |
193 | | (including the terminating '\0'). |
194 | | @param error if not NULL, points to a location where the status of |
195 | | conversion is stored upon return. |
196 | | false successful conversion |
197 | | true the input number is [-,+]infinity or nan. |
198 | | The output string in this case is always '0'. |
199 | | |
200 | | @return number of written characters (excluding terminating '\0') |
201 | | */ |
202 | 0 | size_t my_fcvt(double x, int precision, char *to, bool *error) { |
203 | 0 | return my_fcvt_internal(x, precision, false, to, error); |
204 | 0 | } |
205 | | |
206 | | /// Identical to my_fcvt, except that trailing zeros after the decimal point |
207 | | /// will be removed. |
208 | | size_t my_fcvt_no_trailing_zero(double x, int precision, char *to, |
209 | 0 | bool *error) { |
210 | 0 | return my_fcvt_internal(x, precision, true, to, error); |
211 | 0 | } |
212 | | |
213 | | /** |
214 | | @brief |
215 | | Converts a given floating point number to a zero-terminated string |
216 | | representation using the 'f' format. |
217 | | |
218 | | @details |
219 | | This function is a wrapper around dtoa() to do almost the same as |
220 | | sprintf(to, "%-.*f", precision, x), though the conversion is usually more |
221 | | precise. The only difference is in handling [-,+]infinity and nan values, |
222 | | in which case we print '0\0' to the output string and indicate an overflow. |
223 | | |
224 | | The string always contains the minimum number of digits necessary to |
225 | | reproduce the same binary double value if the string is parsed back to a |
226 | | double value. |
227 | | |
228 | | @param x the input floating point number. |
229 | | @param to pointer to the output buffer. The longest string which |
230 | | my_fcvt() can return is FLOATING_POINT_BUFFER bytes |
231 | | (including the terminating '\0'). |
232 | | @param error if not NULL, points to a location where the status of |
233 | | conversion is stored upon return. |
234 | | false successful conversion |
235 | | true the input number is [-,+]infinity or nan. |
236 | | The output string in this case is always '0'. |
237 | | |
238 | | @return number of written characters (excluding terminating '\0') |
239 | | */ |
240 | 0 | size_t my_fcvt_compact(double x, char *to, bool *error) { |
241 | 0 | return my_fcvt_internal(x, std::numeric_limits<double>::max_digits10, true, |
242 | 0 | to, error); |
243 | 0 | } |
244 | | |
245 | | /** |
246 | | @brief |
247 | | Converts a given floating point number to a zero-terminated string |
248 | | representation with a given field width using the 'e' format |
249 | | (aka scientific notation) or the 'f' one. |
250 | | |
251 | | @details |
252 | | The format is chosen automatically to provide the most number of significant |
253 | | digits (and thus, precision) with a given field width. In many cases, the |
254 | | result is similar to that of sprintf(to, "%g", x) with a few notable |
255 | | differences: |
256 | | - the conversion is usually more precise than C library functions. |
257 | | - there is no 'precision' argument. instead, we specify the number of |
258 | | characters available for conversion (i.e. a field width). |
259 | | - the result never exceeds the specified field width. If the field is too |
260 | | short to contain even a rounded decimal representation, my_gcvt() |
261 | | indicates overflow and truncates the output string to the specified width. |
262 | | - float-type arguments are handled differently than double ones. For a |
263 | | float input number (i.e. when the 'type' argument is MY_GCVT_ARG_FLOAT) |
264 | | we deliberately limit the precision of conversion by FLT_DIG digits to |
265 | | avoid garbage past the significant digits. |
266 | | - unlike sprintf(), in cases where the 'e' format is preferred, we don't |
267 | | zero-pad the exponent to save space for significant digits. The '+' sign |
268 | | for a positive exponent does not appear for the same reason. |
269 | | |
270 | | @param x the input floating point number. |
271 | | @param type is either MY_GCVT_ARG_FLOAT or MY_GCVT_ARG_DOUBLE. |
272 | | Specifies the type of the input number (see notes above). |
273 | | @param width field width in characters. The minimal field width to |
274 | | hold any number representation (albeit rounded) is 7 |
275 | | characters ("-Ne-NNN"). |
276 | | @param to pointer to the output buffer. The result is always |
277 | | zero-terminated, and the longest returned string is thus |
278 | | 'width + 1' bytes. |
279 | | @param error if not NULL, points to a location where the status of |
280 | | conversion is stored upon return. |
281 | | false successful conversion |
282 | | true the input number is [-,+]infinity or nan. |
283 | | The output string in this case is always '0'. |
284 | | @return number of written characters (excluding terminating '\0') |
285 | | |
286 | | @todo |
287 | | Check if it is possible and makes sense to do our own rounding on top of |
288 | | dtoa() instead of calling dtoa() twice in (rare) cases when the resulting |
289 | | string representation does not fit in the specified field width and we want |
290 | | to re-round the input number with fewer significant digits. Examples: |
291 | | |
292 | | my_gcvt(-9e-3, ..., 4, ...); |
293 | | my_gcvt(-9e-3, ..., 2, ...); |
294 | | my_gcvt(1.87e-3, ..., 4, ...); |
295 | | my_gcvt(55, ..., 1, ...); |
296 | | |
297 | | We do our best to minimize such cases by: |
298 | | |
299 | | - passing to dtoa() the field width as the number of significant digits |
300 | | |
301 | | - removing the sign of the number early (and decreasing the width before |
302 | | passing it to dtoa()) |
303 | | |
304 | | - choosing the proper format to preserve the most number of significant |
305 | | digits. |
306 | | */ |
307 | | |
308 | | size_t my_gcvt(double x, my_gcvt_arg_type type, int width, char *to, |
309 | 0 | bool *error) { |
310 | 0 | int decpt, sign, len, exp_len; |
311 | 0 | char *res, *src, *end, *dst = to, *dend = dst + width; |
312 | 0 | char buf[DTOA_BUFF_SIZE]; |
313 | 0 | bool have_space, force_e_format; |
314 | 0 | assert(width > 0 && to != nullptr); |
315 | | |
316 | | /* We want to remove '-' from equations early */ |
317 | 0 | if (x < 0.) width--; |
318 | |
|
319 | 0 | res = |
320 | 0 | dtoa(x, 4, type == MY_GCVT_ARG_DOUBLE ? width : std::min(width, FLT_DIG), |
321 | 0 | &decpt, &sign, &end, buf, sizeof(buf)); |
322 | 0 | if (decpt == DTOA_OVERFLOW) { |
323 | 0 | dtoa_free(res, buf, sizeof(buf)); |
324 | 0 | *to++ = '0'; |
325 | 0 | *to = '\0'; |
326 | 0 | if (error != nullptr) *error = true; |
327 | 0 | return 1; |
328 | 0 | } |
329 | | |
330 | 0 | if (error != nullptr) *error = false; |
331 | |
|
332 | 0 | src = res; |
333 | 0 | len = (int)(end - res); |
334 | | |
335 | | /* |
336 | | Number of digits in the exponent from the 'e' conversion. |
337 | | The sign of the exponent is taken into account separately, we don't need |
338 | | to count it here. |
339 | | */ |
340 | 0 | exp_len = 1 + (decpt >= 101 || decpt <= -99) + (decpt >= 11 || decpt <= -9); |
341 | | |
342 | | /* |
343 | | Do we have enough space for all digits in the 'f' format? |
344 | | Let 'len' be the number of significant digits returned by dtoa, |
345 | | and F be the length of the resulting decimal representation. |
346 | | Consider the following cases: |
347 | | 1. decpt <= 0, i.e. we have "0.NNN" => F = len - decpt + 2 |
348 | | 2. 0 < decpt < len, i.e. we have "NNN.NNN" => F = len + 1 |
349 | | 3. len <= decpt, i.e. we have "NNN00" => F = decpt |
350 | | */ |
351 | 0 | have_space = (decpt <= 0 ? len - decpt + 2 |
352 | 0 | : decpt > 0 && decpt < len ? len + 1 |
353 | 0 | : decpt) <= width; |
354 | | /* |
355 | | The following is true when no significant digits can be placed with the |
356 | | specified field width using the 'f' format, and the 'e' format |
357 | | will not be truncated. |
358 | | */ |
359 | 0 | force_e_format = (decpt <= 0 && width <= 2 - decpt && width >= 3 + exp_len); |
360 | | /* |
361 | | Assume that we don't have enough space to place all significant digits in |
362 | | the 'f' format. We have to choose between the 'e' format and the 'f' one |
363 | | to keep as many significant digits as possible. |
364 | | Let E and F be the lengths of decimal representation in the 'e' and 'f' |
365 | | formats, respectively. We want to use the 'f' format if, and only if F <= E. |
366 | | Consider the following cases: |
367 | | 1. decpt <= 0. |
368 | | F = len - decpt + 2 (see above) |
369 | | E = len + (len > 1) + 1 + 1 (decpt <= -99) + (decpt <= -9) + 1 |
370 | | ("N.NNe-MMM") |
371 | | (F <= E) <=> (len == 1 && decpt >= -1) || (len > 1 && decpt >= -2) |
372 | | We also need to ensure that if the 'f' format is chosen, |
373 | | the field width allows us to place at least one significant digit |
374 | | (i.e. width > 2 - decpt). If not, we prefer the 'e' format. |
375 | | 2. 0 < decpt < len |
376 | | F = len + 1 (see above) |
377 | | E = len + 1 + 1 + ... ("N.NNeMMM") |
378 | | F is always less than E. |
379 | | 3. len <= decpt <= width |
380 | | In this case we have enough space to represent the number in the 'f' |
381 | | format, so we prefer it with some exceptions. |
382 | | 4. width < decpt |
383 | | The number cannot be represented in the 'f' format at all, always use |
384 | | the 'e' 'one. |
385 | | */ |
386 | 0 | if ((have_space || |
387 | | /* |
388 | | Not enough space, let's see if the 'f' format provides the most number |
389 | | of significant digits. |
390 | | */ |
391 | 0 | ((decpt <= width && |
392 | 0 | (decpt >= -1 || (decpt == -2 && (len > 1 || !force_e_format)))) && |
393 | 0 | !force_e_format)) && |
394 | | |
395 | | /* |
396 | | Use the 'e' format in some cases even if we have enough space for the |
397 | | 'f' one. See comment for MAX_DECPT_FOR_F_FORMAT. |
398 | | */ |
399 | 0 | (!have_space || (decpt >= -MAX_DECPT_FOR_F_FORMAT + 1 && |
400 | 0 | (decpt <= MAX_DECPT_FOR_F_FORMAT || len > decpt)))) { |
401 | | /* 'f' format */ |
402 | 0 | int i; |
403 | |
|
404 | 0 | width -= (decpt < len) + (decpt <= 0 ? 1 - decpt : 0); |
405 | | |
406 | | /* Do we have to truncate any digits? */ |
407 | 0 | if (width < len) { |
408 | 0 | if (width < decpt) { |
409 | 0 | if (error != nullptr) *error = true; |
410 | 0 | width = decpt; |
411 | 0 | } |
412 | | |
413 | | /* |
414 | | We want to truncate (len - width) least significant digits after the |
415 | | decimal point. For this we are calling dtoa with mode=5, passing the |
416 | | number of significant digits = (len-decpt) - (len-width) = width-decpt |
417 | | */ |
418 | 0 | dtoa_free(res, buf, sizeof(buf)); |
419 | 0 | res = dtoa(x, 5, width - decpt, &decpt, &sign, &end, buf, sizeof(buf)); |
420 | 0 | src = res; |
421 | 0 | len = (int)(end - res); |
422 | 0 | } |
423 | |
|
424 | 0 | if (len == 0) { |
425 | | /* Underflow. Just print '0' and exit */ |
426 | 0 | *dst++ = '0'; |
427 | 0 | goto end; |
428 | 0 | } |
429 | | |
430 | | /* |
431 | | At this point we are sure we have enough space to put all digits |
432 | | returned by dtoa |
433 | | */ |
434 | 0 | if (sign && dst < dend) *dst++ = '-'; |
435 | 0 | if (decpt <= 0) { |
436 | 0 | if (dst < dend) *dst++ = '0'; |
437 | 0 | if (len > 0 && dst < dend) *dst++ = '.'; |
438 | 0 | for (; decpt < 0 && dst < dend; decpt++) *dst++ = '0'; |
439 | 0 | } |
440 | |
|
441 | 0 | for (i = 1; i <= len && dst < dend; i++) { |
442 | 0 | *dst++ = *src++; |
443 | 0 | if (i == decpt && i < len && dst < dend) *dst++ = '.'; |
444 | 0 | } |
445 | 0 | while (i++ <= decpt && dst < dend) *dst++ = '0'; |
446 | 0 | } else { |
447 | | /* 'e' format */ |
448 | 0 | int decpt_sign = 0; |
449 | |
|
450 | 0 | if (--decpt < 0) { |
451 | 0 | decpt = -decpt; |
452 | 0 | width--; |
453 | 0 | decpt_sign = 1; |
454 | 0 | } |
455 | 0 | width -= 1 + exp_len; /* eNNN */ |
456 | |
|
457 | 0 | if (len > 1) width--; |
458 | |
|
459 | 0 | if (width <= 0) { |
460 | | /* Overflow */ |
461 | 0 | if (error != nullptr) *error = true; |
462 | 0 | width = 0; |
463 | 0 | } |
464 | | |
465 | | /* Do we have to truncate any digits? */ |
466 | 0 | if (width < len) { |
467 | | /* Yes, re-convert with a smaller width */ |
468 | 0 | dtoa_free(res, buf, sizeof(buf)); |
469 | 0 | res = dtoa(x, 4, width, &decpt, &sign, &end, buf, sizeof(buf)); |
470 | 0 | src = res; |
471 | 0 | len = (int)(end - res); |
472 | 0 | if (--decpt < 0) decpt = -decpt; |
473 | 0 | } |
474 | | /* |
475 | | At this point we are sure we have enough space to put all digits |
476 | | returned by dtoa |
477 | | */ |
478 | 0 | if (sign && dst < dend) *dst++ = '-'; |
479 | 0 | if (dst < dend) *dst++ = *src++; |
480 | 0 | if (len > 1 && dst < dend) { |
481 | 0 | *dst++ = '.'; |
482 | 0 | while (src < end && dst < dend) *dst++ = *src++; |
483 | 0 | } |
484 | 0 | if (dst < dend) *dst++ = 'e'; |
485 | 0 | if (decpt_sign && dst < dend) *dst++ = '-'; |
486 | |
|
487 | 0 | if (decpt >= 100 && dst < dend) { |
488 | 0 | *dst++ = decpt / 100 + '0'; |
489 | 0 | decpt %= 100; |
490 | 0 | if (dst < dend) *dst++ = decpt / 10 + '0'; |
491 | 0 | } else if (decpt >= 10 && dst < dend) |
492 | 0 | *dst++ = decpt / 10 + '0'; |
493 | 0 | if (dst < dend) *dst++ = decpt % 10 + '0'; |
494 | 0 | } |
495 | | |
496 | 0 | end: |
497 | 0 | dtoa_free(res, buf, sizeof(buf)); |
498 | 0 | *dst = '\0'; |
499 | |
|
500 | 0 | return dst - to; |
501 | 0 | } |
502 | | |
503 | | /** |
504 | | @brief |
505 | | Converts string to double (string does not have to be zero-terminated) |
506 | | |
507 | | @details |
508 | | This is a wrapper around dtoa's version of strtod(). |
509 | | |
510 | | @param str input string |
511 | | @param end address of a pointer to the first character after the input |
512 | | string. Upon return the pointer is set to point to the first |
513 | | rejected character. |
514 | | @param error Upon return is set to EOVERFLOW in case of underflow or |
515 | | overflow. |
516 | | |
517 | | @return The resulting double value. In case of underflow, 0.0 is |
518 | | returned. In case overflow, signed DBL_MAX is returned. |
519 | | */ |
520 | | |
521 | 0 | double my_strtod(const char *str, const char **end, int *error) { |
522 | 0 | char buf[DTOA_BUFF_SIZE]; |
523 | 0 | double res; |
524 | 0 | assert(end != nullptr && |
525 | 0 | ((str != nullptr && *end != nullptr) || |
526 | 0 | (str == nullptr && *end == nullptr)) && |
527 | 0 | error != nullptr); |
528 | |
|
529 | 0 | res = my_strtod_int(str, end, error, buf, sizeof(buf)); |
530 | 0 | return (*error == 0) ? res : (res < 0 ? -DBL_MAX : DBL_MAX); |
531 | 0 | } |
532 | | |
533 | | /**************************************************************** |
534 | | * |
535 | | * The author of this software is David M. Gay. |
536 | | * |
537 | | * Copyright (c) 1991, 2000, 2001 by Lucent Technologies. |
538 | | * |
539 | | * Permission to use, copy, modify, and distribute this software for any |
540 | | * purpose without fee is hereby granted, provided that this entire notice |
541 | | * is included in all copies of any software which is or includes a copy |
542 | | * or modification of this software and in all copies of the supporting |
543 | | * documentation for such software. |
544 | | * |
545 | | * THIS SOFTWARE IS BEING PROVIDED "AS IS", WITHOUT ANY EXPRESS OR IMPLIED |
546 | | * WARRANTY. IN PARTICULAR, NEITHER THE AUTHOR NOR LUCENT MAKES ANY |
547 | | * REPRESENTATION OR WARRANTY OF ANY KIND CONCERNING THE MERCHANTABILITY |
548 | | * OF THIS SOFTWARE OR ITS FITNESS FOR ANY PARTICULAR PURPOSE. |
549 | | * |
550 | | ***************************************************************/ |
551 | | /* Please send bug reports to David M. Gay (dmg at acm dot org, |
552 | | * with " at " changed at "@" and " dot " changed to "."). */ |
553 | | |
554 | | /* |
555 | | Original copy of the software is located at http://www.netlib.org/fp/dtoa.c |
556 | | It was adjusted to serve MySQL server needs: |
557 | | * strtod() was modified to not expect a zero-terminated string. |
558 | | It now honors 'se' (end of string) argument as the input parameter, |
559 | | not just as the output one. |
560 | | * in dtoa(), in case of overflow/underflow/NaN result string now contains "0"; |
561 | | decpt is set to DTOA_OVERFLOW to indicate overflow. |
562 | | * support for VAX, IBM mainframe and 16-bit hardware removed |
563 | | * we always assume that 64-bit integer type is available |
564 | | * support for Kernigan-Ritchie style headers (pre-ANSI compilers) |
565 | | removed |
566 | | * all gcc warnings ironed out |
567 | | * we always assume multithreaded environment, so we had to change |
568 | | memory allocation procedures to use stack in most cases; |
569 | | malloc is used as the last resort. |
570 | | * pow5mult rewritten to use pre-calculated pow5 list instead of |
571 | | the one generated on the fly. |
572 | | */ |
573 | | |
574 | | /* |
575 | | On a machine with IEEE extended-precision registers, it is |
576 | | necessary to specify double-precision (53-bit) rounding precision |
577 | | before invoking strtod or dtoa. If the machine uses (the equivalent |
578 | | of) Intel 80x87 arithmetic, the call |
579 | | _control87(PC_53, MCW_PC); |
580 | | does this with many compilers. Whether this or another call is |
581 | | appropriate depends on the compiler; for this to work, it may be |
582 | | necessary to #include "float.h" or another system-dependent header |
583 | | file. |
584 | | */ |
585 | | |
586 | | /* |
587 | | #define Honor_FLT_ROUNDS if FLT_ROUNDS can assume the values 2 or 3 |
588 | | and dtoa should round accordingly. |
589 | | #define Check_FLT_ROUNDS if FLT_ROUNDS can assume the values 2 or 3 |
590 | | and Honor_FLT_ROUNDS is not #defined. |
591 | | |
592 | | TODO: check if we can get rid of the above two |
593 | | */ |
594 | | |
595 | | typedef union { |
596 | | double d; |
597 | | uint32_t L[2]; |
598 | | } U; |
599 | | |
600 | | #if defined(WORDS_BIGENDIAN) |
601 | | #define word0(x) (x)->L[0] |
602 | | #define word1(x) (x)->L[1] |
603 | | #else |
604 | 0 | #define word0(x) (x)->L[1] |
605 | 0 | #define word1(x) (x)->L[0] |
606 | | #endif |
607 | | |
608 | 0 | #define dval(x) (x)->d |
609 | | |
610 | | /* #define P DBL_MANT_DIG */ |
611 | | /* Ten_pmax= floor(P*log(2)/log(5)) */ |
612 | | /* Bletch= (highest power of 2 < DBL_MAX_10_EXP) / 16 */ |
613 | | /* Quick_max= floor((P-1)*log(FLT_RADIX)/log(10) - 1) */ |
614 | | /* Int_max= floor(P*log(FLT_RADIX)/log(10) - 1) */ |
615 | | |
616 | 0 | #define Exp_shift 20 |
617 | 0 | #define Exp_shift1 20 |
618 | 0 | #define Exp_msk1 0x100000 |
619 | 0 | #define Exp_mask 0x7ff00000 |
620 | 0 | #define P 53 |
621 | 0 | #define Bias 1023 |
622 | 0 | #define Emin (-1022) |
623 | 0 | #define Exp_1 0x3ff00000 |
624 | 0 | #define Exp_11 0x3ff00000 |
625 | 0 | #define Ebits 11 |
626 | 0 | #define Frac_mask 0xfffff |
627 | 0 | #define Frac_mask1 0xfffff |
628 | 0 | #define Ten_pmax 22 |
629 | 0 | #define Bletch 0x10 |
630 | 0 | #define Bndry_mask 0xfffff |
631 | 0 | #define Bndry_mask1 0xfffff |
632 | 0 | #define LSB 1 |
633 | 0 | #define Sign_bit 0x80000000 |
634 | 0 | #define Log2P 1 |
635 | 0 | #define Tiny1 1 |
636 | 0 | #define Quick_max 14 |
637 | 0 | #define Int_max 14 |
638 | | |
639 | | #ifndef Flt_Rounds |
640 | | #ifdef FLT_ROUNDS |
641 | 0 | #define Flt_Rounds FLT_ROUNDS |
642 | | #else |
643 | | #define Flt_Rounds 1 |
644 | | #endif |
645 | | #endif /*Flt_Rounds*/ |
646 | | |
647 | | #ifdef Honor_FLT_ROUNDS |
648 | | #define Rounding rounding |
649 | | #undef Check_FLT_ROUNDS |
650 | | #define Check_FLT_ROUNDS |
651 | | #endif |
652 | | |
653 | 0 | #define rounded_product(a, b) a *= b |
654 | 0 | #define rounded_quotient(a, b) a /= b |
655 | | |
656 | 0 | #define Big0 (Frac_mask1 | Exp_msk1 * (DBL_MAX_EXP + Bias - 1)) |
657 | 0 | #define Big1 0xffffffff |
658 | 0 | #define FFFFFFFF 0xffffffffUL |
659 | | |
660 | | /* This is tested to be enough for dtoa */ |
661 | | |
662 | 0 | #define Kmax 15 |
663 | | |
664 | | #define Bcopy(x, y) \ |
665 | 0 | memcpy((char *)&x->sign, (char *)&y->sign, \ |
666 | 0 | 2 * sizeof(int) + (y)->wds * sizeof(uint32_t)) |
667 | | |
668 | | /* Arbitrary-length integer */ |
669 | | |
670 | | typedef struct Bigint { |
671 | | union { |
672 | | uint32_t *x; /* points right after this Bigint object */ |
673 | | struct Bigint *next; /* to maintain free lists */ |
674 | | } p; |
675 | | int k; /* 2^k = maxwds */ |
676 | | int maxwds; /* maximum length in 32-bit words */ |
677 | | int sign; /* not zero if number is negative */ |
678 | | int wds; /* current length in 32-bit words */ |
679 | | } Bigint; |
680 | | |
681 | | /* A simple stack-memory based allocator for Bigints */ |
682 | | |
683 | | typedef struct Stack_alloc { |
684 | | char *begin; |
685 | | char *free; |
686 | | char *end; |
687 | | /* |
688 | | Having list of free blocks lets us reduce maximum required amount |
689 | | of memory from ~4000 bytes to < 1680 (tested on x86). |
690 | | */ |
691 | | Bigint *freelist[Kmax + 1]; |
692 | | } Stack_alloc; |
693 | | |
694 | | /* |
695 | | Try to allocate object on stack, and resort to malloc if all |
696 | | stack memory is used. Ensure allocated objects to be aligned by the pointer |
697 | | size in order to not break the alignment rules when storing a pointer to a |
698 | | Bigint. |
699 | | */ |
700 | | |
701 | 0 | static Bigint *Balloc(int k, Stack_alloc *alloc) { |
702 | 0 | Bigint *rv; |
703 | 0 | assert(k <= Kmax); |
704 | 0 | if (k <= Kmax && alloc->freelist[k]) { |
705 | 0 | rv = alloc->freelist[k]; |
706 | 0 | alloc->freelist[k] = rv->p.next; |
707 | 0 | } else { |
708 | 0 | int x, len; |
709 | |
|
710 | 0 | x = 1 << k; |
711 | 0 | len = MY_ALIGN(sizeof(Bigint) + x * sizeof(uint32_t), SIZEOF_CHARP); |
712 | |
|
713 | 0 | if (alloc->free + len <= alloc->end) { |
714 | 0 | rv = (Bigint *)alloc->free; |
715 | 0 | alloc->free += len; |
716 | 0 | } else |
717 | 0 | rv = (Bigint *)malloc(len); |
718 | |
|
719 | 0 | rv->k = k; |
720 | 0 | rv->maxwds = x; |
721 | 0 | } |
722 | 0 | rv->sign = rv->wds = 0; |
723 | 0 | rv->p.x = pointer_cast<uint32_t *>((rv + 1)); |
724 | 0 | return rv; |
725 | 0 | } |
726 | | |
727 | | /* |
728 | | If object was allocated on stack, try putting it to the free |
729 | | list. Otherwise call free(). |
730 | | */ |
731 | | |
732 | 0 | static void Bfree(Bigint *v, Stack_alloc *alloc) { |
733 | 0 | char *gptr = (char *)v; /* generic pointer */ |
734 | 0 | if (gptr < alloc->begin || gptr >= alloc->end) |
735 | 0 | free(gptr); |
736 | 0 | else if (v->k <= Kmax) { |
737 | | /* |
738 | | Maintain free lists only for stack objects: this way we don't |
739 | | have to bother with freeing lists in the end of dtoa; |
740 | | heap should not be used normally anyway. |
741 | | */ |
742 | 0 | v->p.next = alloc->freelist[v->k]; |
743 | 0 | alloc->freelist[v->k] = v; |
744 | 0 | } |
745 | 0 | } |
746 | | |
747 | | /* |
748 | | This is to place return value of dtoa in: tries to use stack |
749 | | as well, but passes by free lists management and just aligns len by |
750 | | the pointer size in order to not break the alignment rules when storing a |
751 | | pointer to a Bigint. |
752 | | */ |
753 | | |
754 | 0 | static char *dtoa_alloc(int i, Stack_alloc *alloc) { |
755 | 0 | char *rv; |
756 | 0 | int const aligned_size = MY_ALIGN(i, SIZEOF_CHARP); |
757 | 0 | if (alloc->free + aligned_size <= alloc->end) { |
758 | 0 | rv = alloc->free; |
759 | 0 | alloc->free += aligned_size; |
760 | 0 | } else |
761 | 0 | rv = static_cast<char *>(malloc(i)); |
762 | 0 | return rv; |
763 | 0 | } |
764 | | |
765 | | /* |
766 | | dtoa_free() must be used to free values s returned by dtoa() |
767 | | This is the counterpart of dtoa_alloc() |
768 | | */ |
769 | | |
770 | 0 | static void dtoa_free(char *gptr, char *buf, size_t buf_size) { |
771 | 0 | if (gptr < buf || gptr >= buf + buf_size) free(gptr); |
772 | 0 | } |
773 | | |
774 | | /* Bigint arithmetic functions */ |
775 | | |
776 | | /* Multiply by m and add a */ |
777 | | |
778 | 0 | static Bigint *multadd(Bigint *b, int m, int a, Stack_alloc *alloc) { |
779 | 0 | uint64_t y = 0; |
780 | 0 | Bigint *b1 = nullptr; |
781 | |
|
782 | 0 | int wds = b->wds; |
783 | 0 | uint32_t *x = b->p.x; |
784 | 0 | int i = 0; |
785 | 0 | uint64_t carry = a; |
786 | 0 | do { |
787 | 0 | y = *x * (uint64_t)m + carry; |
788 | 0 | carry = y >> 32; |
789 | 0 | *x++ = (uint32_t)(y & FFFFFFFF); |
790 | 0 | } while (++i < wds); |
791 | 0 | if (carry) { |
792 | 0 | if (wds >= b->maxwds) { |
793 | 0 | b1 = Balloc(b->k + 1, alloc); |
794 | 0 | Bcopy(b1, b); |
795 | 0 | Bfree(b, alloc); |
796 | 0 | b = b1; |
797 | 0 | } |
798 | 0 | b->p.x[wds++] = (uint32_t)carry; |
799 | 0 | b->wds = wds; |
800 | 0 | } |
801 | 0 | return b; |
802 | 0 | } |
803 | | |
804 | | /** |
805 | | Converts a string to Bigint. |
806 | | |
807 | | Now we have nd0 digits, starting at s, followed by a |
808 | | decimal point, followed by nd-nd0 digits. |
809 | | Unless nd0 == nd, in which case we have a number of the form: |
810 | | ".xxxxxx" or "xxxxxx." |
811 | | |
812 | | @param s Input string, already partially parsed by my_strtod_int(). |
813 | | @param nd0 Number of digits before decimal point. |
814 | | @param nd Total number of digits. |
815 | | @param y9 Pre-computed value of the first nine digits. |
816 | | @param alloc Stack allocator for Bigints. |
817 | | */ |
818 | | static Bigint *s2b(const char *s, int nd0, int nd, uint32_t y9, |
819 | 0 | Stack_alloc *alloc) { |
820 | 0 | Bigint *b; |
821 | 0 | int i, k; |
822 | 0 | int32_t y = 0; |
823 | |
|
824 | 0 | int32_t const x = (nd + 8) / 9; |
825 | 0 | for (k = 0, y = 1; x > y; y <<= 1, k++) |
826 | 0 | ; |
827 | 0 | b = Balloc(k, alloc); |
828 | 0 | b->p.x[0] = y9; |
829 | 0 | b->wds = 1; |
830 | |
|
831 | 0 | i = 9; |
832 | 0 | if (9 < nd0) { |
833 | 0 | s += 9; |
834 | 0 | do b = multadd(b, 10, *s++ - '0', alloc); |
835 | 0 | while (++i < nd0); |
836 | 0 | s++; /* skip '.' */ |
837 | 0 | } else |
838 | 0 | s += 10; |
839 | | /* now do the fractional part */ |
840 | 0 | for (; i < nd; i++) b = multadd(b, 10, *s++ - '0', alloc); |
841 | 0 | return b; |
842 | 0 | } |
843 | | |
844 | 0 | static int hi0bits(uint32_t x) { |
845 | 0 | int k = 0; |
846 | |
|
847 | 0 | if (!(x & 0xffff0000)) { |
848 | 0 | k = 16; |
849 | 0 | x <<= 16; |
850 | 0 | } |
851 | 0 | if (!(x & 0xff000000)) { |
852 | 0 | k += 8; |
853 | 0 | x <<= 8; |
854 | 0 | } |
855 | 0 | if (!(x & 0xf0000000)) { |
856 | 0 | k += 4; |
857 | 0 | x <<= 4; |
858 | 0 | } |
859 | 0 | if (!(x & 0xc0000000)) { |
860 | 0 | k += 2; |
861 | 0 | x <<= 2; |
862 | 0 | } |
863 | 0 | if (!(x & 0x80000000)) { |
864 | 0 | k++; |
865 | 0 | if (!(x & 0x40000000)) return 32; |
866 | 0 | } |
867 | 0 | return k; |
868 | 0 | } |
869 | | |
870 | 0 | static int lo0bits(uint32_t *y) { |
871 | 0 | int k; |
872 | 0 | uint32_t x = *y; |
873 | |
|
874 | 0 | if (x & 7) { |
875 | 0 | if (x & 1) return 0; |
876 | 0 | if (x & 2) { |
877 | 0 | *y = x >> 1; |
878 | 0 | return 1; |
879 | 0 | } |
880 | 0 | *y = x >> 2; |
881 | 0 | return 2; |
882 | 0 | } |
883 | 0 | k = 0; |
884 | 0 | if (!(x & 0xffff)) { |
885 | 0 | k = 16; |
886 | 0 | x >>= 16; |
887 | 0 | } |
888 | 0 | if (!(x & 0xff)) { |
889 | 0 | k += 8; |
890 | 0 | x >>= 8; |
891 | 0 | } |
892 | 0 | if (!(x & 0xf)) { |
893 | 0 | k += 4; |
894 | 0 | x >>= 4; |
895 | 0 | } |
896 | 0 | if (!(x & 0x3)) { |
897 | 0 | k += 2; |
898 | 0 | x >>= 2; |
899 | 0 | } |
900 | 0 | if (!(x & 1)) { |
901 | 0 | k++; |
902 | 0 | x >>= 1; |
903 | 0 | if (!x) return 32; |
904 | 0 | } |
905 | 0 | *y = x; |
906 | 0 | return k; |
907 | 0 | } |
908 | | |
909 | | /* Convert integer to Bigint number */ |
910 | | |
911 | 0 | static Bigint *i2b(int i, Stack_alloc *alloc) { |
912 | 0 | Bigint *b; |
913 | |
|
914 | 0 | b = Balloc(1, alloc); |
915 | 0 | b->p.x[0] = i; |
916 | 0 | b->wds = 1; |
917 | 0 | return b; |
918 | 0 | } |
919 | | |
920 | | /* Multiply two Bigint numbers */ |
921 | | |
922 | 0 | static Bigint *mult(Bigint *a, Bigint *b, Stack_alloc *alloc) { |
923 | 0 | Bigint *c; |
924 | 0 | int k, wa, wb, wc; |
925 | 0 | uint32_t *x = nullptr; |
926 | 0 | uint32_t *xa = nullptr; |
927 | 0 | uint32_t *xc = nullptr; |
928 | 0 | uint32_t y = 0; |
929 | 0 | uint64_t carry = 0; |
930 | 0 | uint64_t z = 0; |
931 | |
|
932 | 0 | if (a->wds < b->wds) { |
933 | 0 | c = a; |
934 | 0 | a = b; |
935 | 0 | b = c; |
936 | 0 | } |
937 | 0 | k = a->k; |
938 | 0 | wa = a->wds; |
939 | 0 | wb = b->wds; |
940 | 0 | wc = wa + wb; |
941 | 0 | if (wc > a->maxwds) k++; |
942 | 0 | c = Balloc(k, alloc); |
943 | 0 | for (x = c->p.x, xa = x + wc; x < xa; x++) *x = 0; |
944 | 0 | xa = a->p.x; |
945 | 0 | uint32_t *xae = xa + wa; |
946 | 0 | uint32_t *xb = b->p.x; |
947 | 0 | uint32_t *xbe = xb + wb; |
948 | 0 | uint32_t *xc0 = c->p.x; |
949 | 0 | for (; xb < xbe; xc0++) { |
950 | 0 | if ((y = *xb++)) { |
951 | 0 | x = xa; |
952 | 0 | xc = xc0; |
953 | 0 | carry = 0; |
954 | 0 | do { |
955 | 0 | z = *x++ * (uint64_t)y + *xc + carry; |
956 | 0 | carry = z >> 32; |
957 | 0 | *xc++ = (uint32_t)(z & FFFFFFFF); |
958 | 0 | } while (x < xae); |
959 | 0 | *xc = (uint32_t)carry; |
960 | 0 | } |
961 | 0 | } |
962 | 0 | for (xc0 = c->p.x, xc = xc0 + wc; wc > 0 && !*--xc; --wc) |
963 | 0 | ; |
964 | 0 | c->wds = wc; |
965 | 0 | return c; |
966 | 0 | } |
967 | | |
968 | | /* |
969 | | Precalculated array of powers of 5: tested to be enough for |
970 | | vasting majority of dtoa_r cases. |
971 | | */ |
972 | | |
973 | | static uint32_t powers5[] = { |
974 | | 625UL, |
975 | | |
976 | | 390625UL, |
977 | | |
978 | | 2264035265UL, 35UL, |
979 | | |
980 | | 2242703233UL, 762134875UL, 1262UL, |
981 | | |
982 | | 3211403009UL, 1849224548UL, 3668416493UL, 3913284084UL, 1593091UL, |
983 | | |
984 | | 781532673UL, 64985353UL, 253049085UL, 594863151UL, 3553621484UL, |
985 | | 3288652808UL, 3167596762UL, 2788392729UL, 3911132675UL, 590UL, |
986 | | |
987 | | 2553183233UL, 3201533787UL, 3638140786UL, 303378311UL, 1809731782UL, |
988 | | 3477761648UL, 3583367183UL, 649228654UL, 2915460784UL, 487929380UL, |
989 | | 1011012442UL, 1677677582UL, 3428152256UL, 1710878487UL, 1438394610UL, |
990 | | 2161952759UL, 4100910556UL, 1608314830UL, 349175UL}; |
991 | | |
992 | | static Bigint p5_a[] = { |
993 | | /* { x } - k - maxwds - sign - wds */ |
994 | | {{powers5}, 1, 1, 0, 1}, {{powers5 + 1}, 1, 1, 0, 1}, |
995 | | {{powers5 + 2}, 1, 2, 0, 2}, {{powers5 + 4}, 2, 3, 0, 3}, |
996 | | {{powers5 + 7}, 3, 5, 0, 5}, {{powers5 + 12}, 4, 10, 0, 10}, |
997 | | {{powers5 + 22}, 5, 19, 0, 19}}; |
998 | | |
999 | 0 | #define P5A_MAX (sizeof(p5_a) / sizeof(*p5_a) - 1) |
1000 | | |
1001 | 0 | static Bigint *pow5mult(Bigint *b, int k, Stack_alloc *alloc) { |
1002 | 0 | Bigint *b1, *p5, *p51 = nullptr; |
1003 | 0 | int i; |
1004 | 0 | static int const p05[3] = {5, 25, 125}; |
1005 | 0 | bool overflow = false; |
1006 | |
|
1007 | 0 | if ((i = k & 3)) b = multadd(b, p05[i - 1], 0, alloc); |
1008 | |
|
1009 | 0 | if (!(k >>= 2)) return b; |
1010 | 0 | p5 = p5_a; |
1011 | 0 | for (;;) { |
1012 | 0 | if (k & 1) { |
1013 | 0 | b1 = mult(b, p5, alloc); |
1014 | 0 | Bfree(b, alloc); |
1015 | 0 | b = b1; |
1016 | 0 | } |
1017 | 0 | if (!(k >>= 1)) break; |
1018 | | /* Calculate next power of 5 */ |
1019 | 0 | if (overflow) { |
1020 | 0 | p51 = mult(p5, p5, alloc); |
1021 | 0 | Bfree(p5, alloc); |
1022 | 0 | p5 = p51; |
1023 | 0 | } else if (p5 < p5_a + P5A_MAX) |
1024 | 0 | ++p5; |
1025 | 0 | else if (p5 == p5_a + P5A_MAX) { |
1026 | 0 | p5 = mult(p5, p5, alloc); |
1027 | 0 | overflow = true; |
1028 | 0 | } |
1029 | 0 | } |
1030 | 0 | if (p51) Bfree(p51, alloc); |
1031 | 0 | return b; |
1032 | 0 | } |
1033 | | |
1034 | 0 | static Bigint *lshift(Bigint *b, int k, Stack_alloc *alloc) { |
1035 | 0 | int i, k1, n, n1; |
1036 | 0 | Bigint *b1; |
1037 | 0 | uint32_t z = 0; |
1038 | |
|
1039 | 0 | n = k >> 5; |
1040 | 0 | k1 = b->k; |
1041 | 0 | n1 = n + b->wds + 1; |
1042 | 0 | for (i = b->maxwds; n1 > i; i <<= 1) k1++; |
1043 | 0 | b1 = Balloc(k1, alloc); |
1044 | 0 | uint32_t *x1 = b1->p.x; |
1045 | 0 | for (i = 0; i < n; i++) *x1++ = 0; |
1046 | 0 | uint32_t *x = b->p.x; |
1047 | 0 | uint32_t *xe = x + b->wds; |
1048 | 0 | if (k &= 0x1f) { |
1049 | 0 | k1 = 32 - k; |
1050 | 0 | z = 0; |
1051 | 0 | do { |
1052 | 0 | *x1++ = *x << k | z; |
1053 | 0 | z = *x++ >> k1; |
1054 | 0 | } while (x < xe); |
1055 | 0 | if ((*x1 = z)) ++n1; |
1056 | 0 | } else |
1057 | 0 | do *x1++ = *x++; |
1058 | 0 | while (x < xe); |
1059 | 0 | b1->wds = n1 - 1; |
1060 | 0 | Bfree(b, alloc); |
1061 | 0 | return b1; |
1062 | 0 | } |
1063 | | |
1064 | 0 | static int cmp(Bigint *a, Bigint *b) { |
1065 | 0 | int i, j; |
1066 | |
|
1067 | 0 | i = a->wds; |
1068 | 0 | j = b->wds; |
1069 | 0 | if (i -= j) return i; |
1070 | 0 | uint32_t *xa0 = a->p.x; |
1071 | 0 | uint32_t *xa = xa0 + j; |
1072 | 0 | uint32_t *xb0 = b->p.x; |
1073 | 0 | uint32_t *xb = xb0 + j; |
1074 | 0 | for (;;) { |
1075 | 0 | if (*--xa != *--xb) return *xa < *xb ? -1 : 1; |
1076 | 0 | if (xa <= xa0) break; |
1077 | 0 | } |
1078 | 0 | return 0; |
1079 | 0 | } |
1080 | | |
1081 | 0 | static Bigint *diff(Bigint *a, Bigint *b, Stack_alloc *alloc) { |
1082 | 0 | Bigint *c = nullptr; |
1083 | 0 | int i, wa, wb; |
1084 | 0 | uint64_t borrow = 0; |
1085 | 0 | uint64_t y = 0; |
1086 | |
|
1087 | 0 | i = cmp(a, b); |
1088 | 0 | if (!i) { |
1089 | 0 | c = Balloc(0, alloc); |
1090 | 0 | c->wds = 1; |
1091 | 0 | c->p.x[0] = 0; |
1092 | 0 | return c; |
1093 | 0 | } |
1094 | 0 | if (i < 0) { |
1095 | 0 | c = a; |
1096 | 0 | a = b; |
1097 | 0 | b = c; |
1098 | 0 | i = 1; |
1099 | 0 | } else |
1100 | 0 | i = 0; |
1101 | 0 | c = Balloc(a->k, alloc); |
1102 | 0 | c->sign = i; |
1103 | 0 | wa = a->wds; |
1104 | 0 | uint32_t *xa = a->p.x; |
1105 | 0 | uint32_t *xae = xa + wa; |
1106 | 0 | wb = b->wds; |
1107 | 0 | uint32_t *xb = b->p.x; |
1108 | 0 | uint32_t *xbe = xb + wb; |
1109 | 0 | uint32_t *xc = c->p.x; |
1110 | 0 | borrow = 0; |
1111 | 0 | do { |
1112 | 0 | y = (uint64_t)*xa++ - *xb++ - borrow; |
1113 | 0 | borrow = y >> 32 & (uint32_t)1; |
1114 | 0 | *xc++ = (uint32_t)(y & FFFFFFFF); |
1115 | 0 | } while (xb < xbe); |
1116 | 0 | while (xa < xae) { |
1117 | 0 | y = *xa++ - borrow; |
1118 | 0 | borrow = y >> 32 & (uint32_t)1; |
1119 | 0 | *xc++ = (uint32_t)(y & FFFFFFFF); |
1120 | 0 | } |
1121 | 0 | while (!*--xc) wa--; |
1122 | 0 | c->wds = wa; |
1123 | 0 | return c; |
1124 | 0 | } |
1125 | | |
1126 | 0 | static double ulp(U *x) { |
1127 | 0 | U u; |
1128 | |
|
1129 | 0 | int32_t const L = (word0(x) & Exp_mask) - (P - 1) * Exp_msk1; |
1130 | 0 | word0(&u) = L; |
1131 | 0 | word1(&u) = 0; |
1132 | 0 | return dval(&u); |
1133 | 0 | } |
1134 | | |
1135 | 0 | static double b2d(Bigint *a, int *e) { |
1136 | 0 | uint32_t z = 0; |
1137 | 0 | int k; |
1138 | 0 | U d; |
1139 | 0 | #define d0 word0(&d) |
1140 | 0 | #define d1 word1(&d) |
1141 | |
|
1142 | 0 | uint32_t *xa0 = a->p.x; |
1143 | 0 | uint32_t *xa = xa0 + a->wds; |
1144 | 0 | uint32_t y = *--xa; |
1145 | 0 | k = hi0bits(y); |
1146 | 0 | *e = 32 - k; |
1147 | 0 | if (k < Ebits) { |
1148 | 0 | d0 = Exp_1 | y >> (Ebits - k); |
1149 | 0 | uint32_t const w = xa > xa0 ? *--xa : 0; |
1150 | 0 | d1 = y << ((32 - Ebits) + k) | w >> (Ebits - k); |
1151 | 0 | goto ret_d; |
1152 | 0 | } |
1153 | 0 | z = xa > xa0 ? *--xa : 0; |
1154 | 0 | if (k -= Ebits) { |
1155 | 0 | d0 = Exp_1 | y << k | z >> (32 - k); |
1156 | 0 | y = xa > xa0 ? *--xa : 0; |
1157 | 0 | d1 = z << k | y >> (32 - k); |
1158 | 0 | } else { |
1159 | 0 | d0 = Exp_1 | y; |
1160 | 0 | d1 = z; |
1161 | 0 | } |
1162 | 0 | ret_d: |
1163 | 0 | #undef d0 |
1164 | 0 | #undef d1 |
1165 | 0 | return dval(&d); |
1166 | 0 | } |
1167 | | |
1168 | 0 | static Bigint *d2b(U *d, int *e, int *bits, Stack_alloc *alloc) { |
1169 | 0 | Bigint *b; |
1170 | 0 | int de, k; |
1171 | 0 | uint32_t y = 0; |
1172 | 0 | uint32_t z = 0; |
1173 | 0 | int i; |
1174 | 0 | #define d0 word0(d) |
1175 | 0 | #define d1 word1(d) |
1176 | |
|
1177 | 0 | b = Balloc(1, alloc); |
1178 | 0 | uint32_t *x = b->p.x; |
1179 | |
|
1180 | 0 | z = d0 & Frac_mask; |
1181 | 0 | d0 &= 0x7fffffff; /* clear sign bit, which we ignore */ |
1182 | 0 | if ((de = (int)(d0 >> Exp_shift))) z |= Exp_msk1; |
1183 | 0 | if ((y = d1)) { |
1184 | 0 | if ((k = lo0bits(&y))) { |
1185 | 0 | x[0] = y | z << (32 - k); |
1186 | 0 | z >>= k; |
1187 | 0 | } else |
1188 | 0 | x[0] = y; |
1189 | 0 | i = b->wds = (x[1] = z) ? 2 : 1; |
1190 | 0 | } else { |
1191 | 0 | k = lo0bits(&z); |
1192 | 0 | x[0] = z; |
1193 | 0 | i = b->wds = 1; |
1194 | 0 | k += 32; |
1195 | 0 | } |
1196 | 0 | if (de) { |
1197 | 0 | *e = de - Bias - (P - 1) + k; |
1198 | 0 | *bits = P - k; |
1199 | 0 | } else { |
1200 | 0 | *e = de - Bias - (P - 1) + 1 + k; |
1201 | 0 | *bits = 32 * i - hi0bits(x[i - 1]); |
1202 | 0 | } |
1203 | 0 | return b; |
1204 | 0 | #undef d0 |
1205 | 0 | #undef d1 |
1206 | 0 | } |
1207 | | |
1208 | 0 | static double ratio(Bigint *a, Bigint *b) { |
1209 | 0 | U da, db; |
1210 | 0 | int k, ka, kb; |
1211 | |
|
1212 | 0 | dval(&da) = b2d(a, &ka); |
1213 | 0 | dval(&db) = b2d(b, &kb); |
1214 | 0 | k = ka - kb + 32 * (a->wds - b->wds); |
1215 | 0 | if (k > 0) |
1216 | 0 | word0(&da) += k * Exp_msk1; |
1217 | 0 | else { |
1218 | 0 | k = -k; |
1219 | 0 | word0(&db) += k * Exp_msk1; |
1220 | 0 | } |
1221 | 0 | return dval(&da) / dval(&db); |
1222 | 0 | } |
1223 | | |
1224 | | static const double tens[] = {1e0, 1e1, 1e2, 1e3, 1e4, 1e5, 1e6, 1e7, |
1225 | | 1e8, 1e9, 1e10, 1e11, 1e12, 1e13, 1e14, 1e15, |
1226 | | 1e16, 1e17, 1e18, 1e19, 1e20, 1e21, 1e22}; |
1227 | | |
1228 | | static const double bigtens[] = {1e16, 1e32, 1e64, 1e128, 1e256}; |
1229 | | static const double tinytens[] = { |
1230 | | 1e-16, 1e-32, 1e-64, 1e-128, |
1231 | | 9007199254740992. * 9007199254740992.e-256 /* = 2^106 * 1e-53 */ |
1232 | | }; |
1233 | | /* |
1234 | | The factor of 2^53 in tinytens[4] helps us avoid setting the underflow |
1235 | | flag unnecessarily. It leads to a song and dance at the end of strtod. |
1236 | | */ |
1237 | 0 | #define Scale_Bit 0x10 |
1238 | 0 | #define n_bigtens 5 |
1239 | | |
1240 | | /* |
1241 | | strtod for IEEE--arithmetic machines. |
1242 | | |
1243 | | This strtod returns a nearest machine number to the input decimal |
1244 | | string (or sets errno to EOVERFLOW). Ties are broken by the IEEE round-even |
1245 | | rule. |
1246 | | |
1247 | | Inspired loosely by William D. Clinger's paper "How to Read Floating |
1248 | | Point Numbers Accurately" [Proc. ACM SIGPLAN '90, pp. 92-101]. |
1249 | | |
1250 | | Modifications: |
1251 | | |
1252 | | 1. We only require IEEE (not IEEE double-extended). |
1253 | | 2. We get by with floating-point arithmetic in a case that |
1254 | | Clinger missed -- when we're computing d * 10^n |
1255 | | for a small integer d and the integer n is not too |
1256 | | much larger than 22 (the maximum integer k for which |
1257 | | we can represent 10^k exactly), we may be able to |
1258 | | compute (d*10^k) * 10^(e-k) with just one roundoff. |
1259 | | 3. Rather than a bit-at-a-time adjustment of the binary |
1260 | | result in the hard case, we use floating-point |
1261 | | arithmetic to determine the adjustment to within |
1262 | | one bit; only in really hard cases do we need to |
1263 | | compute a second residual. |
1264 | | 4. Because of 3., we don't need a large table of powers of 10 |
1265 | | for ten-to-e (just some small tables, e.g. of 10^k |
1266 | | for 0 <= k <= 22). |
1267 | | */ |
1268 | | |
1269 | | static double my_strtod_int(const char *s00, const char **se, int *error, |
1270 | 0 | char *buf, size_t buf_size) { |
1271 | 0 | int scale; |
1272 | 0 | int bb2, bb5, bbe, bd2, bd5, bbbits, bs2, c = 0, dsign, e, e1, esign, i, j, k, |
1273 | 0 | nd, nd0, nf, nz, nz0, sign; |
1274 | 0 | const char *s, *s0, *s1, *end = *se; |
1275 | 0 | double aadj, aadj1; |
1276 | 0 | U aadj2, adj, rv, rv0; |
1277 | 0 | int32_t L = 0; |
1278 | 0 | uint32_t y = 0; |
1279 | 0 | uint32_t z = 0; |
1280 | 0 | Bigint *bb = nullptr, *bb1, *bd = nullptr, *bd0, *bs = nullptr, |
1281 | 0 | *delta = nullptr; |
1282 | | #ifdef Honor_FLT_ROUNDS |
1283 | | int rounding; |
1284 | | #endif |
1285 | 0 | Stack_alloc alloc; |
1286 | |
|
1287 | 0 | *error = 0; |
1288 | |
|
1289 | 0 | alloc.begin = alloc.free = buf; |
1290 | 0 | alloc.end = buf + buf_size; |
1291 | 0 | memset(alloc.freelist, 0, sizeof(alloc.freelist)); |
1292 | |
|
1293 | 0 | sign = nz0 = nz = 0; |
1294 | 0 | dval(&rv) = 0.; |
1295 | 0 | for (s = s00; s < end; s++) switch (*s) { |
1296 | 0 | case '-': |
1297 | 0 | sign = 1; |
1298 | 0 | [[fallthrough]]; |
1299 | 0 | case '+': |
1300 | 0 | s++; |
1301 | 0 | goto break2; |
1302 | 0 | case '\t': |
1303 | 0 | case '\n': |
1304 | 0 | case '\v': |
1305 | 0 | case '\f': |
1306 | 0 | case '\r': |
1307 | 0 | case ' ': |
1308 | 0 | continue; |
1309 | 0 | default: |
1310 | 0 | goto break2; |
1311 | 0 | } |
1312 | 0 | break2: |
1313 | 0 | if (s >= end) goto ret0; |
1314 | | |
1315 | | // Gobble up leading zeros. |
1316 | 0 | if (*s == '0') { |
1317 | 0 | nz0 = 1; |
1318 | 0 | while (++s < end && *s == '0') |
1319 | 0 | ; |
1320 | 0 | if (s >= end) goto ret; |
1321 | 0 | } |
1322 | 0 | s0 = s; |
1323 | 0 | y = z = 0; |
1324 | 0 | for (nd = nf = 0; s < end && (c = *s) >= '0' && c <= '9'; nd++, s++) |
1325 | 0 | if (nd < 9) |
1326 | 0 | y = 10 * y + c - '0'; |
1327 | 0 | else if (nd < 16) |
1328 | 0 | z = 10 * z + c - '0'; |
1329 | 0 | nd0 = nd; |
1330 | 0 | if (s < end && c == '.') { |
1331 | 0 | if (++s < end) c = *s; |
1332 | | // Only leading zeros, now count number of leading zeros after the '.' |
1333 | 0 | if (!nd) { |
1334 | 0 | for (; s < end; ++s) { |
1335 | 0 | c = *s; |
1336 | 0 | if (c != '0') break; |
1337 | 0 | nz++; |
1338 | 0 | } |
1339 | 0 | if (s < end && c > '0' && c <= '9') { |
1340 | 0 | s0 = s; |
1341 | 0 | nf += nz; |
1342 | 0 | nz = 0; |
1343 | 0 | } else |
1344 | 0 | goto dig_done; |
1345 | 0 | } |
1346 | 0 | for (; s < end; ++s) { |
1347 | 0 | c = *s; |
1348 | 0 | if (c < '0' || c > '9') break; |
1349 | | |
1350 | | // We have seen some digits, but not enough of them are non-zero. |
1351 | | // Gobble up all the rest of the digits, and look for exponent. |
1352 | 0 | if (nd > 0 && nz > DBL_MAX_10_EXP) { |
1353 | 0 | continue; |
1354 | 0 | } |
1355 | | |
1356 | | /* |
1357 | | Here we are parsing the fractional part. |
1358 | | We can stop counting digits after a while: the extra digits |
1359 | | will not contribute to the actual result produced by s2b(). |
1360 | | We have to continue scanning, in case there is an exponent part. |
1361 | | */ |
1362 | 0 | if (nd < 2 * DBL_DIG) { |
1363 | 0 | nz++; |
1364 | 0 | if (c -= '0') { |
1365 | 0 | nf += nz; |
1366 | 0 | for (i = 1; i < nz; i++) |
1367 | 0 | if (nd++ < 9) |
1368 | 0 | y *= 10; |
1369 | 0 | else if (nd <= DBL_DIG + 1) |
1370 | 0 | z *= 10; |
1371 | 0 | if (nd++ < 9) |
1372 | 0 | y = 10 * y + c; |
1373 | 0 | else if (nd <= DBL_DIG + 1) |
1374 | 0 | z = 10 * z + c; |
1375 | 0 | nz = 0; |
1376 | 0 | } |
1377 | 0 | } |
1378 | 0 | } |
1379 | 0 | } |
1380 | 0 | dig_done: |
1381 | 0 | e = 0; |
1382 | 0 | if (s < end && (c == 'e' || c == 'E')) { |
1383 | 0 | if (!nd && !nz && !nz0) goto ret0; |
1384 | 0 | s00 = s; |
1385 | 0 | esign = 0; |
1386 | 0 | if (++s < end) switch (c = *s) { |
1387 | 0 | case '-': |
1388 | 0 | esign = 1; |
1389 | 0 | [[fallthrough]]; |
1390 | 0 | case '+': |
1391 | 0 | if (++s < end) c = *s; |
1392 | 0 | } |
1393 | 0 | if (s < end && c >= '0' && c <= '9') { |
1394 | 0 | while (s < end && *s == '0') ++s; // Skip leading zeros in exponent. |
1395 | 0 | if (s < end) c = *s; // First significant digit in exponent, if any. |
1396 | 0 | if (s < end && c > '0' && c <= '9') { |
1397 | 0 | L = c - '0'; |
1398 | 0 | s1 = s; |
1399 | | // Avoid overflow in loop body below. |
1400 | 0 | while (++s < end && (c = *s) >= '0' && c <= '9' && |
1401 | 0 | L < (std::numeric_limits<int32_t>::max() - 255) / 10) { |
1402 | 0 | L = 10 * L + c - '0'; |
1403 | 0 | } |
1404 | 0 | if (s - s1 > 8 || L > 19999) |
1405 | | /* Avoid confusion from exponents |
1406 | | * so large that e might overflow. |
1407 | | */ |
1408 | 0 | e = 19999; /* safe for 16 bit ints */ |
1409 | 0 | else |
1410 | 0 | e = (int)L; |
1411 | 0 | if (esign) e = -e; |
1412 | 0 | } else |
1413 | 0 | e = 0; |
1414 | 0 | } else |
1415 | 0 | s = s00; |
1416 | 0 | } |
1417 | 0 | if (!nd) { |
1418 | 0 | if (!nz && !nz0) { |
1419 | 0 | ret0: |
1420 | 0 | s = s00; |
1421 | 0 | sign = 0; |
1422 | 0 | } |
1423 | 0 | goto ret; |
1424 | 0 | } |
1425 | 0 | e1 = e -= nf; |
1426 | | |
1427 | | /* |
1428 | | Now we have nd0 digits, starting at s0, followed by a |
1429 | | decimal point, followed by nd-nd0 digits. The number we're |
1430 | | after is the integer represented by those digits times |
1431 | | 10**e |
1432 | | */ |
1433 | |
|
1434 | 0 | if (!nd0) nd0 = nd; |
1435 | 0 | k = nd < DBL_DIG + 1 ? nd : DBL_DIG + 1; |
1436 | 0 | dval(&rv) = y; |
1437 | 0 | if (k > 9) { |
1438 | 0 | dval(&rv) = tens[k - 9] * dval(&rv) + z; |
1439 | 0 | } |
1440 | 0 | bd0 = nullptr; |
1441 | 0 | if (nd <= DBL_DIG |
1442 | 0 | #ifndef Honor_FLT_ROUNDS |
1443 | 0 | && Flt_Rounds == 1 |
1444 | 0 | #endif |
1445 | 0 | ) { |
1446 | 0 | if (!e) goto ret; |
1447 | 0 | if (e > 0) { |
1448 | 0 | if (e <= Ten_pmax) { |
1449 | | #ifdef Honor_FLT_ROUNDS |
1450 | | /* round correctly FLT_ROUNDS = 2 or 3 */ |
1451 | | if (sign) { |
1452 | | rv.d = -rv.d; |
1453 | | sign = 0; |
1454 | | } |
1455 | | #endif |
1456 | 0 | /* rv = */ rounded_product(dval(&rv), tens[e]); |
1457 | 0 | goto ret; |
1458 | 0 | } |
1459 | 0 | i = DBL_DIG - nd; |
1460 | 0 | if (e <= Ten_pmax + i) { |
1461 | | /* |
1462 | | A fancier test would sometimes let us do |
1463 | | this for larger i values. |
1464 | | */ |
1465 | | #ifdef Honor_FLT_ROUNDS |
1466 | | /* round correctly FLT_ROUNDS = 2 or 3 */ |
1467 | | if (sign) { |
1468 | | rv.d = -rv.d; |
1469 | | sign = 0; |
1470 | | } |
1471 | | #endif |
1472 | 0 | e -= i; |
1473 | 0 | dval(&rv) *= tens[i]; |
1474 | 0 | /* rv = */ rounded_product(dval(&rv), tens[e]); |
1475 | 0 | goto ret; |
1476 | 0 | } |
1477 | 0 | } |
1478 | 0 | #ifndef Inaccurate_Divide |
1479 | 0 | else if (e >= -Ten_pmax) { |
1480 | | #ifdef Honor_FLT_ROUNDS |
1481 | | /* round correctly FLT_ROUNDS = 2 or 3 */ |
1482 | | if (sign) { |
1483 | | rv.d = -rv.d; |
1484 | | sign = 0; |
1485 | | } |
1486 | | #endif |
1487 | 0 | /* rv = */ rounded_quotient(dval(&rv), tens[-e]); |
1488 | 0 | goto ret; |
1489 | 0 | } |
1490 | 0 | #endif |
1491 | 0 | } |
1492 | 0 | e1 += nd - k; |
1493 | |
|
1494 | 0 | scale = 0; |
1495 | | #ifdef Honor_FLT_ROUNDS |
1496 | | if ((rounding = Flt_Rounds) >= 2) { |
1497 | | if (sign) |
1498 | | rounding = rounding == 2 ? 0 : 2; |
1499 | | else if (rounding != 2) |
1500 | | rounding = 0; |
1501 | | } |
1502 | | #endif |
1503 | | |
1504 | | /* Get starting approximation = rv * 10**e1 */ |
1505 | |
|
1506 | 0 | if (e1 > 0) { |
1507 | 0 | if ((i = e1 & 15)) dval(&rv) *= tens[i]; |
1508 | 0 | if (e1 &= ~15) { |
1509 | 0 | if (e1 > DBL_MAX_10_EXP) { |
1510 | 0 | ovfl: |
1511 | 0 | *error = EOVERFLOW; |
1512 | | /* Can't trust HUGE_VAL */ |
1513 | | #ifdef Honor_FLT_ROUNDS |
1514 | | switch (rounding) { |
1515 | | case 0: /* toward 0 */ |
1516 | | case 3: /* toward -infinity */ |
1517 | | word0(&rv) = Big0; |
1518 | | word1(&rv) = Big1; |
1519 | | break; |
1520 | | default: |
1521 | | word0(&rv) = Exp_mask; |
1522 | | word1(&rv) = 0; |
1523 | | } |
1524 | | #else /*Honor_FLT_ROUNDS*/ |
1525 | 0 | word0(&rv) = Exp_mask; |
1526 | 0 | word1(&rv) = 0; |
1527 | 0 | #endif /*Honor_FLT_ROUNDS*/ |
1528 | 0 | if (bd0) goto retfree; |
1529 | 0 | goto ret; |
1530 | 0 | } |
1531 | 0 | e1 >>= 4; |
1532 | 0 | for (j = 0; e1 > 1; j++, e1 >>= 1) |
1533 | 0 | if (e1 & 1) dval(&rv) *= bigtens[j]; |
1534 | | /* The last multiplication could overflow. */ |
1535 | 0 | word0(&rv) -= P * Exp_msk1; |
1536 | 0 | dval(&rv) *= bigtens[j]; |
1537 | 0 | if ((z = word0(&rv) & Exp_mask) > Exp_msk1 * (DBL_MAX_EXP + Bias - P)) |
1538 | 0 | goto ovfl; |
1539 | 0 | if (z > Exp_msk1 * (DBL_MAX_EXP + Bias - 1 - P)) { |
1540 | | /* set to largest number (Can't trust DBL_MAX) */ |
1541 | 0 | word0(&rv) = Big0; |
1542 | 0 | word1(&rv) = Big1; |
1543 | 0 | } else |
1544 | 0 | word0(&rv) += P * Exp_msk1; |
1545 | 0 | } |
1546 | 0 | } else if (e1 < 0) { |
1547 | 0 | e1 = -e1; |
1548 | 0 | if ((i = e1 & 15)) dval(&rv) /= tens[i]; |
1549 | 0 | if ((e1 >>= 4)) { |
1550 | 0 | if (e1 >= 1 << n_bigtens) goto undfl; |
1551 | 0 | if (e1 & Scale_Bit) scale = 2 * P; |
1552 | 0 | for (j = 0; e1 > 0; j++, e1 >>= 1) |
1553 | 0 | if (e1 & 1) dval(&rv) *= tinytens[j]; |
1554 | 0 | if (scale && |
1555 | 0 | (j = 2 * P + 1 - ((word0(&rv) & Exp_mask) >> Exp_shift)) > 0) { |
1556 | | /* scaled rv is denormal; zap j low bits */ |
1557 | 0 | if (j >= 32) { |
1558 | 0 | word1(&rv) = 0; |
1559 | 0 | if (j >= 53) |
1560 | 0 | word0(&rv) = (P + 2) * Exp_msk1; |
1561 | 0 | else |
1562 | 0 | word0(&rv) &= 0xffffffff << (j - 32); |
1563 | 0 | } else |
1564 | 0 | word1(&rv) &= 0xffffffff << j; |
1565 | 0 | } |
1566 | 0 | if (!dval(&rv)) { |
1567 | 0 | undfl: |
1568 | 0 | dval(&rv) = 0.; |
1569 | 0 | if (bd0) goto retfree; |
1570 | 0 | goto ret; |
1571 | 0 | } |
1572 | 0 | } |
1573 | 0 | } |
1574 | | |
1575 | | /* Now the hard part -- adjusting rv to the correct value.*/ |
1576 | | |
1577 | | /* Put digits into bd: true value = bd * 10^e */ |
1578 | | |
1579 | 0 | bd0 = s2b(s0, nd0, nd, y, &alloc); |
1580 | |
|
1581 | 0 | for (;;) { |
1582 | 0 | bd = Balloc(bd0->k, &alloc); |
1583 | 0 | Bcopy(bd, bd0); |
1584 | 0 | bb = d2b(&rv, &bbe, &bbbits, &alloc); /* rv = bb * 2^bbe */ |
1585 | 0 | bs = i2b(1, &alloc); |
1586 | |
|
1587 | 0 | if (e >= 0) { |
1588 | 0 | bb2 = bb5 = 0; |
1589 | 0 | bd2 = bd5 = e; |
1590 | 0 | } else { |
1591 | 0 | bb2 = bb5 = -e; |
1592 | 0 | bd2 = bd5 = 0; |
1593 | 0 | } |
1594 | 0 | if (bbe >= 0) |
1595 | 0 | bb2 += bbe; |
1596 | 0 | else |
1597 | 0 | bd2 -= bbe; |
1598 | 0 | bs2 = bb2; |
1599 | | #ifdef Honor_FLT_ROUNDS |
1600 | | if (rounding != 1) bs2++; |
1601 | | #endif |
1602 | 0 | j = bbe - scale; |
1603 | 0 | i = j + bbbits - 1; /* logb(rv) */ |
1604 | 0 | if (i < Emin) /* denormal */ |
1605 | 0 | j += P - Emin; |
1606 | 0 | else |
1607 | 0 | j = P + 1 - bbbits; |
1608 | 0 | bb2 += j; |
1609 | 0 | bd2 += j; |
1610 | 0 | bd2 += scale; |
1611 | 0 | i = bb2 < bd2 ? bb2 : bd2; |
1612 | 0 | if (i > bs2) i = bs2; |
1613 | 0 | if (i > 0) { |
1614 | 0 | bb2 -= i; |
1615 | 0 | bd2 -= i; |
1616 | 0 | bs2 -= i; |
1617 | 0 | } |
1618 | 0 | if (bb5 > 0) { |
1619 | 0 | bs = pow5mult(bs, bb5, &alloc); |
1620 | 0 | bb1 = mult(bs, bb, &alloc); |
1621 | 0 | Bfree(bb, &alloc); |
1622 | 0 | bb = bb1; |
1623 | 0 | } |
1624 | 0 | if (bb2 > 0) bb = lshift(bb, bb2, &alloc); |
1625 | 0 | if (bd5 > 0) bd = pow5mult(bd, bd5, &alloc); |
1626 | 0 | if (bd2 > 0) bd = lshift(bd, bd2, &alloc); |
1627 | 0 | if (bs2 > 0) bs = lshift(bs, bs2, &alloc); |
1628 | 0 | delta = diff(bb, bd, &alloc); |
1629 | 0 | dsign = delta->sign; |
1630 | 0 | delta->sign = 0; |
1631 | 0 | i = cmp(delta, bs); |
1632 | | #ifdef Honor_FLT_ROUNDS |
1633 | | if (rounding != 1) { |
1634 | | if (i < 0) { |
1635 | | /* Error is less than an ulp */ |
1636 | | if (!delta->p.x[0] && delta->wds <= 1) { |
1637 | | /* exact */ |
1638 | | break; |
1639 | | } |
1640 | | if (rounding) { |
1641 | | if (dsign) { |
1642 | | adj.d = 1.; |
1643 | | goto apply_adj; |
1644 | | } |
1645 | | } else if (!dsign) { |
1646 | | adj.d = -1.; |
1647 | | if (!word1(&rv) && !(word0(&rv) & Frac_mask)) { |
1648 | | y = word0(&rv) & Exp_mask; |
1649 | | if (!scale || y > 2 * P * Exp_msk1) { |
1650 | | delta = lshift(delta, Log2P, &alloc); |
1651 | | if (cmp(delta, bs) <= 0) adj.d = -0.5; |
1652 | | } |
1653 | | } |
1654 | | apply_adj: |
1655 | | if (scale && (y = word0(&rv) & Exp_mask) <= 2 * P * Exp_msk1) |
1656 | | word0(&adj) += (2 * P + 1) * Exp_msk1 - y; |
1657 | | dval(&rv) += adj.d * ulp(&rv); |
1658 | | } |
1659 | | break; |
1660 | | } |
1661 | | adj.d = ratio(delta, bs); |
1662 | | if (adj.d < 1.) adj.d = 1.; |
1663 | | if (adj.d <= 0x7ffffffe) { |
1664 | | /* adj = rounding ? ceil(adj) : floor(adj); */ |
1665 | | y = adj.d; |
1666 | | if (y != adj.d) { |
1667 | | if (!((rounding >> 1) ^ dsign)) y++; |
1668 | | adj.d = y; |
1669 | | } |
1670 | | } |
1671 | | if (scale && (y = word0(&rv) & Exp_mask) <= 2 * P * Exp_msk1) |
1672 | | word0(&adj) += (2 * P + 1) * Exp_msk1 - y; |
1673 | | adj.d *= ulp(&rv); |
1674 | | if (dsign) |
1675 | | dval(&rv) += adj.d; |
1676 | | else |
1677 | | dval(&rv) -= adj.d; |
1678 | | goto cont; |
1679 | | } |
1680 | | #endif /*Honor_FLT_ROUNDS*/ |
1681 | |
|
1682 | 0 | if (i < 0) { |
1683 | | /* |
1684 | | Error is less than half an ulp -- check for special case of mantissa |
1685 | | a power of two. |
1686 | | */ |
1687 | 0 | if (dsign || word1(&rv) || word0(&rv) & Bndry_mask || |
1688 | 0 | (word0(&rv) & Exp_mask) <= (2 * P + 1) * Exp_msk1) { |
1689 | 0 | break; |
1690 | 0 | } |
1691 | 0 | if (!delta->p.x[0] && delta->wds <= 1) { |
1692 | | /* exact result */ |
1693 | 0 | break; |
1694 | 0 | } |
1695 | 0 | delta = lshift(delta, Log2P, &alloc); |
1696 | 0 | if (cmp(delta, bs) > 0) goto drop_down; |
1697 | 0 | break; |
1698 | 0 | } |
1699 | 0 | if (i == 0) { |
1700 | | /* exactly half-way between */ |
1701 | 0 | if (dsign) { |
1702 | 0 | if ((word0(&rv) & Bndry_mask1) == Bndry_mask1 && |
1703 | 0 | word1(&rv) == |
1704 | 0 | ((scale && (y = word0(&rv) & Exp_mask) <= 2 * P * Exp_msk1) |
1705 | 0 | ? (0xffffffff & |
1706 | 0 | (0xffffffff << (2 * P + 1 - (y >> Exp_shift)))) |
1707 | 0 | : 0xffffffff)) { |
1708 | | /*boundary case -- increment exponent*/ |
1709 | 0 | word0(&rv) = (word0(&rv) & Exp_mask) + Exp_msk1; |
1710 | 0 | word1(&rv) = 0; |
1711 | 0 | dsign = 0; |
1712 | 0 | break; |
1713 | 0 | } |
1714 | 0 | } else if (!(word0(&rv) & Bndry_mask) && !word1(&rv)) { |
1715 | 0 | drop_down: |
1716 | | /* boundary case -- decrement exponent */ |
1717 | 0 | if (scale) { |
1718 | 0 | L = word0(&rv) & Exp_mask; |
1719 | 0 | if (L <= (2 * P + 1) * Exp_msk1) { |
1720 | 0 | if (L > (P + 2) * Exp_msk1) /* round even ==> accept rv */ |
1721 | 0 | break; |
1722 | | /* rv = smallest denormal */ |
1723 | 0 | goto undfl; |
1724 | 0 | } |
1725 | 0 | } |
1726 | 0 | L = (word0(&rv) & Exp_mask) - Exp_msk1; |
1727 | 0 | word0(&rv) = L | Bndry_mask1; |
1728 | 0 | word1(&rv) = 0xffffffff; |
1729 | 0 | break; |
1730 | 0 | } |
1731 | 0 | if (!(word1(&rv) & LSB)) break; |
1732 | 0 | if (dsign) |
1733 | 0 | dval(&rv) += ulp(&rv); |
1734 | 0 | else { |
1735 | 0 | dval(&rv) -= ulp(&rv); |
1736 | 0 | if (!dval(&rv)) goto undfl; |
1737 | 0 | } |
1738 | 0 | dsign = 1 - dsign; |
1739 | 0 | break; |
1740 | 0 | } |
1741 | 0 | if ((aadj = ratio(delta, bs)) <= 2.) { |
1742 | 0 | if (dsign) |
1743 | 0 | aadj = aadj1 = 1.; |
1744 | 0 | else if (word1(&rv) || word0(&rv) & Bndry_mask) { |
1745 | 0 | if (word1(&rv) == Tiny1 && !word0(&rv)) goto undfl; |
1746 | 0 | aadj = 1.; |
1747 | 0 | aadj1 = -1.; |
1748 | 0 | } else { |
1749 | | /* special case -- power of FLT_RADIX to be rounded down... */ |
1750 | 0 | if (aadj < 2. / FLT_RADIX) |
1751 | 0 | aadj = 1. / FLT_RADIX; |
1752 | 0 | else |
1753 | 0 | aadj *= 0.5; |
1754 | 0 | aadj1 = -aadj; |
1755 | 0 | } |
1756 | 0 | } else { |
1757 | 0 | aadj *= 0.5; |
1758 | 0 | aadj1 = dsign ? aadj : -aadj; |
1759 | | #ifdef Check_FLT_ROUNDS |
1760 | | switch (Rounding) { |
1761 | | case 2: /* towards +infinity */ |
1762 | | aadj1 -= 0.5; |
1763 | | break; |
1764 | | case 0: /* towards 0 */ |
1765 | | case 3: /* towards -infinity */ |
1766 | | aadj1 += 0.5; |
1767 | | } |
1768 | | #else |
1769 | 0 | if (Flt_Rounds == 0) aadj1 += 0.5; |
1770 | 0 | #endif /*Check_FLT_ROUNDS*/ |
1771 | 0 | } |
1772 | 0 | y = word0(&rv) & Exp_mask; |
1773 | | |
1774 | | /* Check for overflow */ |
1775 | |
|
1776 | 0 | if (y == Exp_msk1 * (DBL_MAX_EXP + Bias - 1)) { |
1777 | 0 | dval(&rv0) = dval(&rv); |
1778 | 0 | word0(&rv) -= P * Exp_msk1; |
1779 | 0 | adj.d = aadj1 * ulp(&rv); |
1780 | 0 | dval(&rv) += adj.d; |
1781 | 0 | if ((word0(&rv) & Exp_mask) >= Exp_msk1 * (DBL_MAX_EXP + Bias - P)) { |
1782 | 0 | if (word0(&rv0) == Big0 && word1(&rv0) == Big1) goto ovfl; |
1783 | 0 | word0(&rv) = Big0; |
1784 | 0 | word1(&rv) = Big1; |
1785 | 0 | goto cont; |
1786 | 0 | } else |
1787 | 0 | word0(&rv) += P * Exp_msk1; |
1788 | 0 | } else { |
1789 | 0 | if (scale && y <= 2 * P * Exp_msk1) { |
1790 | 0 | if (aadj <= 0x7fffffff) { |
1791 | 0 | if ((z = (uint32_t)aadj) <= 0) z = 1; |
1792 | 0 | aadj = z; |
1793 | 0 | aadj1 = dsign ? aadj : -aadj; |
1794 | 0 | } |
1795 | 0 | dval(&aadj2) = aadj1; |
1796 | 0 | word0(&aadj2) += (2 * P + 1) * Exp_msk1 - y; |
1797 | 0 | aadj1 = dval(&aadj2); |
1798 | 0 | adj.d = aadj1 * ulp(&rv); |
1799 | 0 | dval(&rv) += adj.d; |
1800 | 0 | if (rv.d == 0.) goto undfl; |
1801 | 0 | } else { |
1802 | 0 | adj.d = aadj1 * ulp(&rv); |
1803 | 0 | dval(&rv) += adj.d; |
1804 | 0 | } |
1805 | 0 | } |
1806 | 0 | z = word0(&rv) & Exp_mask; |
1807 | 0 | if (!scale) |
1808 | 0 | if (y == z) { |
1809 | | /* Can we stop now? */ |
1810 | 0 | L = (int32_t)aadj; |
1811 | 0 | aadj -= L; |
1812 | | /* The tolerances below are conservative. */ |
1813 | 0 | if (dsign || word1(&rv) || word0(&rv) & Bndry_mask) { |
1814 | 0 | if (aadj < .4999999 || aadj > .5000001) break; |
1815 | 0 | } else if (aadj < .4999999 / FLT_RADIX) |
1816 | 0 | break; |
1817 | 0 | } |
1818 | 0 | cont: |
1819 | 0 | Bfree(bb, &alloc); |
1820 | 0 | Bfree(bd, &alloc); |
1821 | 0 | Bfree(bs, &alloc); |
1822 | 0 | Bfree(delta, &alloc); |
1823 | 0 | } |
1824 | 0 | if (scale) { |
1825 | 0 | word0(&rv0) = Exp_1 - 2 * P * Exp_msk1; |
1826 | 0 | word1(&rv0) = 0; |
1827 | 0 | dval(&rv) *= dval(&rv0); |
1828 | 0 | } |
1829 | 0 | retfree: |
1830 | 0 | Bfree(bb, &alloc); |
1831 | 0 | Bfree(bd, &alloc); |
1832 | 0 | Bfree(bs, &alloc); |
1833 | 0 | Bfree(bd0, &alloc); |
1834 | 0 | Bfree(delta, &alloc); |
1835 | 0 | ret: |
1836 | 0 | *se = s; |
1837 | 0 | return sign ? -dval(&rv) : dval(&rv); |
1838 | 0 | } |
1839 | | |
1840 | 0 | static int quorem(Bigint *b, Bigint *S) { |
1841 | 0 | int n; |
1842 | 0 | uint32_t *bx = nullptr; |
1843 | 0 | uint32_t *bxe = nullptr; |
1844 | 0 | uint32_t *sx = nullptr; |
1845 | 0 | uint32_t *sxe = nullptr; |
1846 | 0 | uint64_t borrow = 0; |
1847 | 0 | uint64_t carry = 0; |
1848 | 0 | uint64_t y = 0; |
1849 | 0 | uint64_t ys = 0; |
1850 | |
|
1851 | 0 | n = S->wds; |
1852 | 0 | if (b->wds < n) return 0; |
1853 | 0 | sx = S->p.x; |
1854 | 0 | sxe = sx + --n; |
1855 | 0 | bx = b->p.x; |
1856 | 0 | bxe = bx + n; |
1857 | 0 | uint32_t q = *bxe / (*sxe + 1); /* ensure q <= true quotient */ |
1858 | 0 | if (q) { |
1859 | 0 | borrow = 0; |
1860 | 0 | carry = 0; |
1861 | 0 | do { |
1862 | 0 | ys = *sx++ * (uint64_t)q + carry; |
1863 | 0 | carry = ys >> 32; |
1864 | 0 | y = *bx - (ys & FFFFFFFF) - borrow; |
1865 | 0 | borrow = y >> 32 & (uint32_t)1; |
1866 | 0 | *bx++ = (uint32_t)(y & FFFFFFFF); |
1867 | 0 | } while (sx <= sxe); |
1868 | 0 | if (!*bxe) { |
1869 | 0 | bx = b->p.x; |
1870 | 0 | while (--bxe > bx && !*bxe) --n; |
1871 | 0 | b->wds = n; |
1872 | 0 | } |
1873 | 0 | } |
1874 | 0 | if (cmp(b, S) >= 0) { |
1875 | 0 | q++; |
1876 | 0 | borrow = 0; |
1877 | 0 | carry = 0; |
1878 | 0 | bx = b->p.x; |
1879 | 0 | sx = S->p.x; |
1880 | 0 | do { |
1881 | 0 | ys = *sx++ + carry; |
1882 | 0 | carry = ys >> 32; |
1883 | 0 | y = *bx - (ys & FFFFFFFF) - borrow; |
1884 | 0 | borrow = y >> 32 & (uint32_t)1; |
1885 | 0 | *bx++ = (uint32_t)(y & FFFFFFFF); |
1886 | 0 | } while (sx <= sxe); |
1887 | 0 | bx = b->p.x; |
1888 | 0 | bxe = bx + n; |
1889 | 0 | if (!*bxe) { |
1890 | 0 | while (--bxe > bx && !*bxe) --n; |
1891 | 0 | b->wds = n; |
1892 | 0 | } |
1893 | 0 | } |
1894 | 0 | return q; |
1895 | 0 | } |
1896 | | |
1897 | | /* |
1898 | | dtoa for IEEE arithmetic (dmg): convert double to ASCII string. |
1899 | | |
1900 | | Inspired by "How to Print Floating-Point Numbers Accurately" by |
1901 | | Guy L. Steele, Jr. and Jon L. White [Proc. ACM SIGPLAN '90, pp. 112-126]. |
1902 | | |
1903 | | Modifications: |
1904 | | 1. Rather than iterating, we use a simple numeric overestimate |
1905 | | to determine k= floor(log10(d)). We scale relevant |
1906 | | quantities using O(log2(k)) rather than O(k) multiplications. |
1907 | | 2. For some modes > 2 (corresponding to ecvt and fcvt), we don't |
1908 | | try to generate digits strictly left to right. Instead, we |
1909 | | compute with fewer bits and propagate the carry if necessary |
1910 | | when rounding the final digit up. This is often faster. |
1911 | | 3. Under the assumption that input will be rounded nearest, |
1912 | | mode 0 renders 1e23 as 1e23 rather than 9.999999999999999e22. |
1913 | | That is, we allow equality in stopping tests when the |
1914 | | round-nearest rule will give the same floating-point value |
1915 | | as would satisfaction of the stopping test with strict |
1916 | | inequality. |
1917 | | 4. We remove common factors of powers of 2 from relevant |
1918 | | quantities. |
1919 | | 5. When converting floating-point integers less than 1e16, |
1920 | | we use floating-point arithmetic rather than resorting |
1921 | | to multiple-precision integers. |
1922 | | 6. When asked to produce fewer than 15 digits, we first try |
1923 | | to get by with floating-point arithmetic; we resort to |
1924 | | multiple-precision integer arithmetic only if we cannot |
1925 | | guarantee that the floating-point calculation has given |
1926 | | the correctly rounded result. For k requested digits and |
1927 | | "uniformly" distributed input, the probability is |
1928 | | something like 10^(k-15) that we must resort to the int32_t |
1929 | | calculation. |
1930 | | */ |
1931 | | |
1932 | | static char *dtoa(double dd, int mode, int ndigits, int *decpt, int *sign, |
1933 | 0 | char **rve, char *buf, size_t buf_size) { |
1934 | | /* |
1935 | | Arguments ndigits, decpt, sign are similar to those |
1936 | | of ecvt and fcvt; trailing zeros are suppressed from |
1937 | | the returned string. If not null, *rve is set to point |
1938 | | to the end of the return value. If d is +-Infinity or NaN, |
1939 | | then *decpt is set to DTOA_OVERFLOW. |
1940 | | |
1941 | | mode: |
1942 | | 0 ==> shortest string that yields d when read in |
1943 | | and rounded to nearest. |
1944 | | 1 ==> like 0, but with Steele & White stopping rule; |
1945 | | e.g. with IEEE P754 arithmetic , mode 0 gives |
1946 | | 1e23 whereas mode 1 gives 9.999999999999999e22. |
1947 | | 2 ==> max(1,ndigits) significant digits. This gives a |
1948 | | return value similar to that of ecvt, except |
1949 | | that trailing zeros are suppressed. |
1950 | | 3 ==> through ndigits past the decimal point. This |
1951 | | gives a return value similar to that from fcvt, |
1952 | | except that trailing zeros are suppressed, and |
1953 | | ndigits can be negative. |
1954 | | 4,5 ==> similar to 2 and 3, respectively, but (in |
1955 | | round-nearest mode) with the tests of mode 0 to |
1956 | | possibly return a shorter string that rounds to d. |
1957 | | With IEEE arithmetic and compilation with |
1958 | | -DHonor_FLT_ROUNDS, modes 4 and 5 behave the same |
1959 | | as modes 2 and 3 when FLT_ROUNDS != 1. |
1960 | | 6-9 ==> Debugging modes similar to mode - 4: don't try |
1961 | | fast floating-point estimate (if applicable). |
1962 | | |
1963 | | Values of mode other than 0-9 are treated as mode 0. |
1964 | | |
1965 | | Sufficient space is allocated to the return value |
1966 | | to hold the suppressed trailing zeros. |
1967 | | */ |
1968 | |
|
1969 | 0 | int bbits, b2, b5, be, dig, i, ieps, ilim = 0, ilim0, ilim1 = 0, j, j1, k, k0, |
1970 | 0 | k_check, leftright, m2, m5, s2, s5, |
1971 | 0 | spec_case, try_quick; |
1972 | 0 | int32_t L = 0; |
1973 | 0 | int denorm; |
1974 | 0 | uint32_t x = 0; |
1975 | 0 | Bigint *b, *b1, *delta, *mlo = nullptr, *mhi, *S; |
1976 | 0 | U d2, eps, u; |
1977 | 0 | double ds; |
1978 | 0 | char *s, *s0; |
1979 | | #ifdef Honor_FLT_ROUNDS |
1980 | | int rounding; |
1981 | | #endif |
1982 | 0 | Stack_alloc alloc; |
1983 | |
|
1984 | 0 | alloc.begin = alloc.free = buf; |
1985 | 0 | alloc.end = buf + buf_size; |
1986 | 0 | memset(alloc.freelist, 0, sizeof(alloc.freelist)); |
1987 | |
|
1988 | 0 | u.d = dd; |
1989 | 0 | if (word0(&u) & Sign_bit) { |
1990 | | /* set sign for everything, including 0's and NaNs */ |
1991 | 0 | *sign = 1; |
1992 | 0 | word0(&u) &= ~Sign_bit; /* clear sign bit */ |
1993 | 0 | } else |
1994 | 0 | *sign = 0; |
1995 | | |
1996 | | /* If infinity, set decpt to DTOA_OVERFLOW, if 0 set it to 1 */ |
1997 | 0 | if (((word0(&u) & Exp_mask) == Exp_mask && (*decpt = DTOA_OVERFLOW)) || |
1998 | 0 | (!dval(&u) && (*decpt = 1))) { |
1999 | | /* Infinity, NaN, 0 */ |
2000 | 0 | char *res = (char *)dtoa_alloc(2, &alloc); |
2001 | 0 | res[0] = '0'; |
2002 | 0 | res[1] = '\0'; |
2003 | 0 | if (rve) *rve = res + 1; |
2004 | 0 | return res; |
2005 | 0 | } |
2006 | | |
2007 | | #ifdef Honor_FLT_ROUNDS |
2008 | | if ((rounding = Flt_Rounds) >= 2) { |
2009 | | if (*sign) |
2010 | | rounding = rounding == 2 ? 0 : 2; |
2011 | | else if (rounding != 2) |
2012 | | rounding = 0; |
2013 | | } |
2014 | | #endif |
2015 | | |
2016 | 0 | b = d2b(&u, &be, &bbits, &alloc); |
2017 | 0 | if ((i = (int)(word0(&u) >> Exp_shift1 & (Exp_mask >> Exp_shift1)))) { |
2018 | 0 | dval(&d2) = dval(&u); |
2019 | 0 | word0(&d2) &= Frac_mask1; |
2020 | 0 | word0(&d2) |= Exp_11; |
2021 | | |
2022 | | /* |
2023 | | log(x) ~=~ log(1.5) + (x-1.5)/1.5 |
2024 | | log10(x) = log(x) / log(10) |
2025 | | ~=~ log(1.5)/log(10) + (x-1.5)/(1.5*log(10)) |
2026 | | log10(d)= (i-Bias)*log(2)/log(10) + log10(d2) |
2027 | | |
2028 | | This suggests computing an approximation k to log10(d) by |
2029 | | |
2030 | | k= (i - Bias)*0.301029995663981 |
2031 | | + ( (d2-1.5)*0.289529654602168 + 0.176091259055681 ); |
2032 | | |
2033 | | We want k to be too large rather than too small. |
2034 | | The error in the first-order Taylor series approximation |
2035 | | is in our favor, so we just round up the constant enough |
2036 | | to compensate for any error in the multiplication of |
2037 | | (i - Bias) by 0.301029995663981; since |i - Bias| <= 1077, |
2038 | | and 1077 * 0.30103 * 2^-52 ~=~ 7.2e-14, |
2039 | | adding 1e-13 to the constant term more than suffices. |
2040 | | Hence we adjust the constant term to 0.1760912590558. |
2041 | | (We could get a more accurate k by invoking log10, |
2042 | | but this is probably not worthwhile.) |
2043 | | */ |
2044 | |
|
2045 | 0 | i -= Bias; |
2046 | 0 | denorm = 0; |
2047 | 0 | } else { |
2048 | | /* d is denormalized */ |
2049 | |
|
2050 | 0 | i = bbits + be + (Bias + (P - 1) - 1); |
2051 | 0 | x = i > 32 ? word0(&u) << (64 - i) | word1(&u) >> (i - 32) |
2052 | 0 | : word1(&u) << (32 - i); |
2053 | 0 | dval(&d2) = x; |
2054 | 0 | word0(&d2) -= 31 * Exp_msk1; /* adjust exponent */ |
2055 | 0 | i -= (Bias + (P - 1) - 1) + 1; |
2056 | 0 | denorm = 1; |
2057 | 0 | } |
2058 | 0 | ds = (dval(&d2) - 1.5) * 0.289529654602168 + 0.1760912590558 + |
2059 | 0 | i * 0.301029995663981; |
2060 | 0 | k = (int)ds; |
2061 | 0 | if (ds < 0. && ds != k) k--; /* want k= floor(ds) */ |
2062 | 0 | k_check = 1; |
2063 | 0 | if (k >= 0 && k <= Ten_pmax) { |
2064 | 0 | if (dval(&u) < tens[k]) k--; |
2065 | 0 | k_check = 0; |
2066 | 0 | } |
2067 | 0 | j = bbits - i - 1; |
2068 | 0 | if (j >= 0) { |
2069 | 0 | b2 = 0; |
2070 | 0 | s2 = j; |
2071 | 0 | } else { |
2072 | 0 | b2 = -j; |
2073 | 0 | s2 = 0; |
2074 | 0 | } |
2075 | 0 | if (k >= 0) { |
2076 | 0 | b5 = 0; |
2077 | 0 | s5 = k; |
2078 | 0 | s2 += k; |
2079 | 0 | } else { |
2080 | 0 | b2 -= k; |
2081 | 0 | b5 = -k; |
2082 | 0 | s5 = 0; |
2083 | 0 | } |
2084 | 0 | if (mode < 0 || mode > 9) mode = 0; |
2085 | |
|
2086 | | #ifdef Check_FLT_ROUNDS |
2087 | | try_quick = Rounding == 1; |
2088 | | #else |
2089 | 0 | try_quick = 1; |
2090 | 0 | #endif |
2091 | |
|
2092 | 0 | if (mode > 5) { |
2093 | 0 | mode -= 4; |
2094 | 0 | try_quick = 0; |
2095 | 0 | } |
2096 | 0 | leftright = 1; |
2097 | 0 | switch (mode) { |
2098 | 0 | case 0: |
2099 | 0 | case 1: |
2100 | 0 | ilim = ilim1 = -1; |
2101 | 0 | i = 18; |
2102 | 0 | ndigits = 0; |
2103 | 0 | break; |
2104 | 0 | case 2: |
2105 | 0 | leftright = 0; |
2106 | 0 | [[fallthrough]]; |
2107 | 0 | case 4: |
2108 | 0 | if (ndigits <= 0) ndigits = 1; |
2109 | 0 | ilim = ilim1 = i = ndigits; |
2110 | 0 | break; |
2111 | 0 | case 3: |
2112 | 0 | leftright = 0; |
2113 | 0 | [[fallthrough]]; |
2114 | 0 | case 5: |
2115 | 0 | i = ndigits + k + 1; |
2116 | 0 | ilim = i; |
2117 | 0 | ilim1 = i - 1; |
2118 | 0 | if (i <= 0) i = 1; |
2119 | 0 | } |
2120 | 0 | s = s0 = dtoa_alloc(i, &alloc); |
2121 | |
|
2122 | | #ifdef Honor_FLT_ROUNDS |
2123 | | if (mode > 1 && rounding != 1) leftright = 0; |
2124 | | #endif |
2125 | |
|
2126 | 0 | if (ilim >= 0 && ilim <= Quick_max && try_quick) { |
2127 | | /* Try to get by with floating-point arithmetic. */ |
2128 | 0 | i = 0; |
2129 | 0 | dval(&d2) = dval(&u); |
2130 | 0 | k0 = k; |
2131 | 0 | ilim0 = ilim; |
2132 | 0 | ieps = 2; /* conservative */ |
2133 | 0 | if (k > 0) { |
2134 | 0 | ds = tens[k & 0xf]; |
2135 | 0 | j = k >> 4; |
2136 | 0 | if (j & Bletch) { |
2137 | | /* prevent overflows */ |
2138 | 0 | j &= Bletch - 1; |
2139 | 0 | dval(&u) /= bigtens[n_bigtens - 1]; |
2140 | 0 | ieps++; |
2141 | 0 | } |
2142 | 0 | for (; j; j >>= 1, i++) { |
2143 | 0 | if (j & 1) { |
2144 | 0 | ieps++; |
2145 | 0 | ds *= bigtens[i]; |
2146 | 0 | } |
2147 | 0 | } |
2148 | 0 | dval(&u) /= ds; |
2149 | 0 | } else if ((j1 = -k)) { |
2150 | 0 | dval(&u) *= tens[j1 & 0xf]; |
2151 | 0 | for (j = j1 >> 4; j; j >>= 1, i++) { |
2152 | 0 | if (j & 1) { |
2153 | 0 | ieps++; |
2154 | 0 | dval(&u) *= bigtens[i]; |
2155 | 0 | } |
2156 | 0 | } |
2157 | 0 | } |
2158 | 0 | if (k_check && dval(&u) < 1. && ilim > 0) { |
2159 | 0 | if (ilim1 <= 0) goto fast_failed; |
2160 | 0 | ilim = ilim1; |
2161 | 0 | k--; |
2162 | 0 | dval(&u) *= 10.; |
2163 | 0 | ieps++; |
2164 | 0 | } |
2165 | 0 | dval(&eps) = ieps * dval(&u) + 7.; |
2166 | 0 | word0(&eps) -= (P - 1) * Exp_msk1; |
2167 | 0 | if (ilim == 0) { |
2168 | 0 | S = mhi = nullptr; |
2169 | 0 | dval(&u) -= 5.; |
2170 | 0 | if (dval(&u) > dval(&eps)) goto one_digit; |
2171 | 0 | if (dval(&u) < -dval(&eps)) goto no_digits; |
2172 | 0 | goto fast_failed; |
2173 | 0 | } |
2174 | 0 | if (leftright) { |
2175 | | /* Use Steele & White method of only generating digits needed. */ |
2176 | 0 | dval(&eps) = 0.5 / tens[ilim - 1] - dval(&eps); |
2177 | 0 | for (i = 0;;) { |
2178 | 0 | L = (int32_t)dval(&u); |
2179 | 0 | dval(&u) -= L; |
2180 | 0 | *s++ = '0' + (int)L; |
2181 | 0 | if (dval(&u) < dval(&eps)) goto ret1; |
2182 | 0 | if (1. - dval(&u) < dval(&eps)) goto bump_up; |
2183 | 0 | if (++i >= ilim) break; |
2184 | 0 | dval(&eps) *= 10.; |
2185 | 0 | dval(&u) *= 10.; |
2186 | 0 | } |
2187 | 0 | } else { |
2188 | | /* Generate ilim digits, then fix them up. */ |
2189 | 0 | dval(&eps) *= tens[ilim - 1]; |
2190 | 0 | for (i = 1;; i++, dval(&u) *= 10.) { |
2191 | 0 | L = (int32_t)(dval(&u)); |
2192 | 0 | if (!(dval(&u) -= L)) ilim = i; |
2193 | 0 | *s++ = '0' + (int)L; |
2194 | 0 | if (i == ilim) { |
2195 | 0 | if (dval(&u) > 0.5 + dval(&eps)) |
2196 | 0 | goto bump_up; |
2197 | 0 | else if (dval(&u) < 0.5 - dval(&eps)) { |
2198 | 0 | while (*--s == '0') |
2199 | 0 | ; |
2200 | 0 | s++; |
2201 | 0 | goto ret1; |
2202 | 0 | } |
2203 | 0 | break; |
2204 | 0 | } |
2205 | 0 | } |
2206 | 0 | } |
2207 | 0 | fast_failed: |
2208 | 0 | s = s0; |
2209 | 0 | dval(&u) = dval(&d2); |
2210 | 0 | k = k0; |
2211 | 0 | ilim = ilim0; |
2212 | 0 | } |
2213 | | |
2214 | | /* Do we have a "small" integer? */ |
2215 | | |
2216 | 0 | if (be >= 0 && k <= Int_max) { |
2217 | | /* Yes. */ |
2218 | 0 | ds = tens[k]; |
2219 | 0 | if (ndigits < 0 && ilim <= 0) { |
2220 | 0 | S = mhi = nullptr; |
2221 | 0 | if (ilim < 0 || dval(&u) <= 5 * ds) goto no_digits; |
2222 | 0 | goto one_digit; |
2223 | 0 | } |
2224 | 0 | for (i = 1;; i++, dval(&u) *= 10.) { |
2225 | 0 | L = (int32_t)(dval(&u) / ds); |
2226 | 0 | dval(&u) -= L * ds; |
2227 | | #ifdef Check_FLT_ROUNDS |
2228 | | /* If FLT_ROUNDS == 2, L will usually be high by 1 */ |
2229 | | if (dval(&u) < 0) { |
2230 | | L--; |
2231 | | dval(&u) += ds; |
2232 | | } |
2233 | | #endif |
2234 | 0 | *s++ = '0' + (int)L; |
2235 | 0 | if (!dval(&u)) { |
2236 | 0 | break; |
2237 | 0 | } |
2238 | 0 | if (i == ilim) { |
2239 | | #ifdef Honor_FLT_ROUNDS |
2240 | | if (mode > 1) { |
2241 | | switch (rounding) { |
2242 | | case 0: |
2243 | | goto ret1; |
2244 | | case 2: |
2245 | | goto bump_up; |
2246 | | } |
2247 | | } |
2248 | | #endif |
2249 | 0 | dval(&u) += dval(&u); |
2250 | 0 | if (dval(&u) > ds || (dval(&u) == ds && L & 1)) { |
2251 | 0 | bump_up: |
2252 | 0 | while (*--s == '9') |
2253 | 0 | if (s == s0) { |
2254 | 0 | k++; |
2255 | 0 | *s = '0'; |
2256 | 0 | break; |
2257 | 0 | } |
2258 | 0 | ++*s++; |
2259 | 0 | } |
2260 | 0 | break; |
2261 | 0 | } |
2262 | 0 | } |
2263 | 0 | goto ret1; |
2264 | 0 | } |
2265 | | |
2266 | 0 | m2 = b2; |
2267 | 0 | m5 = b5; |
2268 | 0 | mhi = mlo = nullptr; |
2269 | 0 | if (leftright) { |
2270 | 0 | i = denorm ? be + (Bias + (P - 1) - 1 + 1) : 1 + P - bbits; |
2271 | 0 | b2 += i; |
2272 | 0 | s2 += i; |
2273 | 0 | mhi = i2b(1, &alloc); |
2274 | 0 | } |
2275 | 0 | if (m2 > 0 && s2 > 0) { |
2276 | 0 | i = m2 < s2 ? m2 : s2; |
2277 | 0 | b2 -= i; |
2278 | 0 | m2 -= i; |
2279 | 0 | s2 -= i; |
2280 | 0 | } |
2281 | 0 | if (b5 > 0) { |
2282 | 0 | if (leftright) { |
2283 | 0 | if (m5 > 0) { |
2284 | 0 | mhi = pow5mult(mhi, m5, &alloc); |
2285 | 0 | b1 = mult(mhi, b, &alloc); |
2286 | 0 | Bfree(b, &alloc); |
2287 | 0 | b = b1; |
2288 | 0 | } |
2289 | 0 | if ((j = b5 - m5)) b = pow5mult(b, j, &alloc); |
2290 | 0 | } else |
2291 | 0 | b = pow5mult(b, b5, &alloc); |
2292 | 0 | } |
2293 | 0 | S = i2b(1, &alloc); |
2294 | 0 | if (s5 > 0) S = pow5mult(S, s5, &alloc); |
2295 | | |
2296 | | /* Check for special case that d is a normalized power of 2. */ |
2297 | |
|
2298 | 0 | spec_case = 0; |
2299 | 0 | if ((mode < 2 || leftright) |
2300 | | #ifdef Honor_FLT_ROUNDS |
2301 | | && rounding == 1 |
2302 | | #endif |
2303 | 0 | ) { |
2304 | 0 | if (!word1(&u) && !(word0(&u) & Bndry_mask) && |
2305 | 0 | word0(&u) & (Exp_mask & ~Exp_msk1)) { |
2306 | | /* The special case */ |
2307 | 0 | b2 += Log2P; |
2308 | 0 | s2 += Log2P; |
2309 | 0 | spec_case = 1; |
2310 | 0 | } |
2311 | 0 | } |
2312 | | |
2313 | | /* |
2314 | | Arrange for convenient computation of quotients: |
2315 | | shift left if necessary so divisor has 4 leading 0 bits. |
2316 | | |
2317 | | Perhaps we should just compute leading 28 bits of S once |
2318 | | a nd for all and pass them and a shift to quorem, so it |
2319 | | can do shifts and ors to compute the numerator for q. |
2320 | | */ |
2321 | 0 | if ((i = ((s5 ? 32 - hi0bits(S->p.x[S->wds - 1]) : 1) + s2) & 0x1f)) |
2322 | 0 | i = 32 - i; |
2323 | 0 | if (i > 4) { |
2324 | 0 | i -= 4; |
2325 | 0 | b2 += i; |
2326 | 0 | m2 += i; |
2327 | 0 | s2 += i; |
2328 | 0 | } else if (i < 4) { |
2329 | 0 | i += 28; |
2330 | 0 | b2 += i; |
2331 | 0 | m2 += i; |
2332 | 0 | s2 += i; |
2333 | 0 | } |
2334 | 0 | if (b2 > 0) b = lshift(b, b2, &alloc); |
2335 | 0 | if (s2 > 0) S = lshift(S, s2, &alloc); |
2336 | 0 | if (k_check) { |
2337 | 0 | if (cmp(b, S) < 0) { |
2338 | 0 | k--; |
2339 | | /* we botched the k estimate */ |
2340 | 0 | b = multadd(b, 10, 0, &alloc); |
2341 | 0 | if (leftright) mhi = multadd(mhi, 10, 0, &alloc); |
2342 | 0 | ilim = ilim1; |
2343 | 0 | } |
2344 | 0 | } |
2345 | 0 | if (ilim <= 0 && (mode == 3 || mode == 5)) { |
2346 | 0 | if (ilim < 0 || cmp(b, S = multadd(S, 5, 0, &alloc)) <= 0) { |
2347 | | /* no digits, fcvt style */ |
2348 | 0 | no_digits: |
2349 | 0 | k = -1 - ndigits; |
2350 | 0 | goto ret; |
2351 | 0 | } |
2352 | 0 | one_digit: |
2353 | 0 | *s++ = '1'; |
2354 | 0 | k++; |
2355 | 0 | goto ret; |
2356 | 0 | } |
2357 | 0 | if (leftright) { |
2358 | 0 | if (m2 > 0) mhi = lshift(mhi, m2, &alloc); |
2359 | | |
2360 | | /* |
2361 | | Compute mlo -- check for special case that d is a normalized power of 2. |
2362 | | */ |
2363 | |
|
2364 | 0 | mlo = mhi; |
2365 | 0 | if (spec_case) { |
2366 | 0 | mhi = Balloc(mhi->k, &alloc); |
2367 | 0 | Bcopy(mhi, mlo); |
2368 | 0 | mhi = lshift(mhi, Log2P, &alloc); |
2369 | 0 | } |
2370 | |
|
2371 | 0 | for (i = 1;; i++) { |
2372 | 0 | dig = quorem(b, S) + '0'; |
2373 | | /* Do we yet have the shortest decimal string that will round to d? */ |
2374 | 0 | j = cmp(b, mlo); |
2375 | 0 | delta = diff(S, mhi, &alloc); |
2376 | 0 | j1 = delta->sign ? 1 : cmp(b, delta); |
2377 | 0 | Bfree(delta, &alloc); |
2378 | 0 | if (j1 == 0 && mode != 1 && !(word1(&u) & 1) |
2379 | | #ifdef Honor_FLT_ROUNDS |
2380 | | && rounding >= 1 |
2381 | | #endif |
2382 | 0 | ) { |
2383 | 0 | if (dig == '9') goto round_9_up; |
2384 | 0 | if (j > 0) dig++; |
2385 | 0 | *s++ = dig; |
2386 | 0 | goto ret; |
2387 | 0 | } |
2388 | 0 | if (j < 0 || (j == 0 && mode != 1 && !(word1(&u) & 1))) { |
2389 | 0 | if (!b->p.x[0] && b->wds <= 1) { |
2390 | 0 | goto accept_dig; |
2391 | 0 | } |
2392 | | #ifdef Honor_FLT_ROUNDS |
2393 | | if (mode > 1) switch (rounding) { |
2394 | | case 0: |
2395 | | goto accept_dig; |
2396 | | case 2: |
2397 | | goto keep_dig; |
2398 | | } |
2399 | | #endif /*Honor_FLT_ROUNDS*/ |
2400 | 0 | if (j1 > 0) { |
2401 | 0 | b = lshift(b, 1, &alloc); |
2402 | 0 | j1 = cmp(b, S); |
2403 | 0 | if ((j1 > 0 || (j1 == 0 && dig & 1)) && dig++ == '9') goto round_9_up; |
2404 | 0 | } |
2405 | 0 | accept_dig: |
2406 | 0 | *s++ = dig; |
2407 | 0 | goto ret; |
2408 | 0 | } |
2409 | 0 | if (j1 > 0) { |
2410 | | #ifdef Honor_FLT_ROUNDS |
2411 | | if (!rounding) goto accept_dig; |
2412 | | #endif |
2413 | 0 | if (dig == '9') { /* possible if i == 1 */ |
2414 | 0 | round_9_up: |
2415 | 0 | *s++ = '9'; |
2416 | 0 | goto roundoff; |
2417 | 0 | } |
2418 | 0 | *s++ = dig + 1; |
2419 | 0 | goto ret; |
2420 | 0 | } |
2421 | | #ifdef Honor_FLT_ROUNDS |
2422 | | keep_dig: |
2423 | | #endif |
2424 | 0 | *s++ = dig; |
2425 | 0 | if (i == ilim) break; |
2426 | 0 | b = multadd(b, 10, 0, &alloc); |
2427 | 0 | if (mlo == mhi) |
2428 | 0 | mlo = mhi = multadd(mhi, 10, 0, &alloc); |
2429 | 0 | else { |
2430 | 0 | mlo = multadd(mlo, 10, 0, &alloc); |
2431 | 0 | mhi = multadd(mhi, 10, 0, &alloc); |
2432 | 0 | } |
2433 | 0 | } |
2434 | 0 | } else |
2435 | 0 | for (i = 1;; i++) { |
2436 | 0 | *s++ = dig = quorem(b, S) + '0'; |
2437 | 0 | if (!b->p.x[0] && b->wds <= 1) { |
2438 | 0 | goto ret; |
2439 | 0 | } |
2440 | 0 | if (i >= ilim) break; |
2441 | 0 | b = multadd(b, 10, 0, &alloc); |
2442 | 0 | } |
2443 | | |
2444 | | /* Round off last digit */ |
2445 | | |
2446 | | #ifdef Honor_FLT_ROUNDS |
2447 | | switch (rounding) { |
2448 | | case 0: |
2449 | | goto trimzeros; |
2450 | | case 2: |
2451 | | goto roundoff; |
2452 | | } |
2453 | | #endif |
2454 | 0 | b = lshift(b, 1, &alloc); |
2455 | 0 | j = cmp(b, S); |
2456 | 0 | if (j > 0 || (j == 0 && dig & 1)) { |
2457 | 0 | roundoff: |
2458 | 0 | while (*--s == '9') |
2459 | 0 | if (s == s0) { |
2460 | 0 | k++; |
2461 | 0 | *s++ = '1'; |
2462 | 0 | goto ret; |
2463 | 0 | } |
2464 | 0 | ++*s++; |
2465 | 0 | } else { |
2466 | | #ifdef Honor_FLT_ROUNDS |
2467 | | trimzeros: |
2468 | | #endif |
2469 | 0 | while (*--s == '0') |
2470 | 0 | ; |
2471 | 0 | s++; |
2472 | 0 | } |
2473 | 0 | ret: |
2474 | 0 | Bfree(S, &alloc); |
2475 | 0 | if (mhi) { |
2476 | 0 | if (mlo && mlo != mhi) Bfree(mlo, &alloc); |
2477 | 0 | Bfree(mhi, &alloc); |
2478 | 0 | } |
2479 | 0 | ret1: |
2480 | 0 | Bfree(b, &alloc); |
2481 | 0 | *s = 0; |
2482 | 0 | *decpt = k + 1; |
2483 | 0 | if (rve) *rve = s; |
2484 | 0 | return s0; |
2485 | 0 | } |