1 | /* Copyright (c) 2007, 2012, Oracle and/or its affiliates. All rights reserved. |
2 | Copyright (c) 2017, MariaDB Corporation. |
3 | |
4 | This library is free software; you can redistribute it and/or |
5 | modify it under the terms of the GNU Library General Public |
6 | License as published by the Free Software Foundation; version 2 |
7 | of the License. |
8 | |
9 | This program is distributed in the hope that it will be useful, |
10 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
11 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
12 | GNU General Public License for more details. |
13 | |
14 | You should have received a copy of the GNU General Public License |
15 | along with this program; if not, write to the Free Software |
16 | Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA */ |
17 | |
18 | /**************************************************************** |
19 | |
20 | This file incorporates work covered by the following copyright and |
21 | permission notice: |
22 | |
23 | The author of this software is David M. Gay. |
24 | |
25 | Copyright (c) 1991, 2000, 2001 by Lucent Technologies. |
26 | |
27 | Permission to use, copy, modify, and distribute this software for any |
28 | purpose without fee is hereby granted, provided that this entire notice |
29 | is included in all copies of any software which is or includes a copy |
30 | or modification of this software and in all copies of the supporting |
31 | documentation for such software. |
32 | |
33 | THIS SOFTWARE IS BEING PROVIDED "AS IS", WITHOUT ANY EXPRESS OR IMPLIED |
34 | WARRANTY. IN PARTICULAR, NEITHER THE AUTHOR NOR LUCENT MAKES ANY |
35 | REPRESENTATION OR WARRANTY OF ANY KIND CONCERNING THE MERCHANTABILITY |
36 | OF THIS SOFTWARE OR ITS FITNESS FOR ANY PARTICULAR PURPOSE. |
37 | |
38 | ***************************************************************/ |
39 | |
40 | #include "strings_def.h" |
41 | #include <my_base.h> /* for EOVERFLOW on Windows */ |
42 | |
43 | /** |
44 | Appears to suffice to not call malloc() in most cases. |
45 | @todo |
46 | see if it is possible to get rid of malloc(). |
47 | this constant is sufficient to avoid malloc() on all inputs I have tried. |
48 | */ |
49 | #define DTOA_BUFF_SIZE (460 * sizeof(void *)) |
50 | |
51 | /* Magic value returned by dtoa() to indicate overflow */ |
52 | #define DTOA_OVERFLOW 9999 |
53 | |
54 | static double my_strtod_int(const char *, char **, int *, char *, size_t); |
55 | static char *dtoa(double, int, int, int *, int *, char **, char *, size_t); |
56 | static void dtoa_free(char *, char *, size_t); |
57 | |
58 | /** |
59 | @brief |
60 | Converts a given floating point number to a zero-terminated string |
61 | representation using the 'f' format. |
62 | |
63 | @details |
64 | This function is a wrapper around dtoa() to do the same as |
65 | sprintf(to, "%-.*f", precision, x), though the conversion is usually more |
66 | precise. The only difference is in handling [-,+]infinity and nan values, |
67 | in which case we print '0\0' to the output string and indicate an overflow. |
68 | |
69 | @param x the input floating point number. |
70 | @param precision the number of digits after the decimal point. |
71 | All properties of sprintf() apply: |
72 | - if the number of significant digits after the decimal |
73 | point is less than precision, the resulting string is |
74 | right-padded with zeros |
75 | - if the precision is 0, no decimal point appears |
76 | - if a decimal point appears, at least one digit appears |
77 | before it |
78 | @param to pointer to the output buffer. The longest string which |
79 | my_fcvt() can return is FLOATING_POINT_BUFFER bytes |
80 | (including the terminating '\0'). |
81 | @param error if not NULL, points to a location where the status of |
82 | conversion is stored upon return. |
83 | FALSE successful conversion |
84 | TRUE the input number is [-,+]infinity or nan. |
85 | The output string in this case is always '0'. |
86 | @return number of written characters (excluding terminating '\0') |
87 | */ |
88 | |
89 | size_t my_fcvt(double x, int precision, char *to, my_bool *error) |
90 | { |
91 | int decpt, sign, len, i; |
92 | char *res, *src, *end, *dst= to; |
93 | char buf[DTOA_BUFF_SIZE]; |
94 | DBUG_ASSERT(precision >= 0 && precision < DECIMAL_NOT_SPECIFIED && to != NULL); |
95 | |
96 | res= dtoa(x, 5, precision, &decpt, &sign, &end, buf, sizeof(buf)); |
97 | |
98 | if (decpt == DTOA_OVERFLOW) |
99 | { |
100 | dtoa_free(res, buf, sizeof(buf)); |
101 | *to++= '0'; |
102 | *to= '\0'; |
103 | if (error != NULL) |
104 | *error= TRUE; |
105 | return 1; |
106 | } |
107 | |
108 | src= res; |
109 | len= (int)(end - src); |
110 | |
111 | if (sign) |
112 | *dst++= '-'; |
113 | |
114 | if (decpt <= 0) |
115 | { |
116 | *dst++= '0'; |
117 | *dst++= '.'; |
118 | for (i= decpt; i < 0; i++) |
119 | *dst++= '0'; |
120 | } |
121 | |
122 | for (i= 1; i <= len; i++) |
123 | { |
124 | *dst++= *src++; |
125 | if (i == decpt && i < len) |
126 | *dst++= '.'; |
127 | } |
128 | while (i++ <= decpt) |
129 | *dst++= '0'; |
130 | |
131 | if (precision > 0) |
132 | { |
133 | if (len <= decpt) |
134 | *dst++= '.'; |
135 | |
136 | for (i= precision - MY_MAX(0, (len - decpt)); i > 0; i--) |
137 | *dst++= '0'; |
138 | } |
139 | |
140 | *dst= '\0'; |
141 | if (error != NULL) |
142 | *error= FALSE; |
143 | |
144 | dtoa_free(res, buf, sizeof(buf)); |
145 | |
146 | return dst - to; |
147 | } |
148 | |
149 | /** |
150 | @brief |
151 | Converts a given floating point number to a zero-terminated string |
152 | representation with a given field width using the 'e' format |
153 | (aka scientific notation) or the 'f' one. |
154 | |
155 | @details |
156 | The format is chosen automatically to provide the most number of significant |
157 | digits (and thus, precision) with a given field width. In many cases, the |
158 | result is similar to that of sprintf(to, "%g", x) with a few notable |
159 | differences: |
160 | - the conversion is usually more precise than C library functions. |
161 | - there is no 'precision' argument. instead, we specify the number of |
162 | characters available for conversion (i.e. a field width). |
163 | - the result never exceeds the specified field width. If the field is too |
164 | short to contain even a rounded decimal representation, my_gcvt() |
165 | indicates overflow and truncates the output string to the specified width. |
166 | - float-type arguments are handled differently than double ones. For a |
167 | float input number (i.e. when the 'type' argument is MY_GCVT_ARG_FLOAT) |
168 | we deliberately limit the precision of conversion by FLT_DIG digits to |
169 | avoid garbage past the significant digits. |
170 | - unlike sprintf(), in cases where the 'e' format is preferred, we don't |
171 | zero-pad the exponent to save space for significant digits. The '+' sign |
172 | for a positive exponent does not appear for the same reason. |
173 | |
174 | @param x the input floating point number. |
175 | @param type is either MY_GCVT_ARG_FLOAT or MY_GCVT_ARG_DOUBLE. |
176 | Specifies the type of the input number (see notes above). |
177 | @param width field width in characters. The minimal field width to |
178 | hold any number representation (albeit rounded) is 7 |
179 | characters ("-Ne-NNN"). |
180 | @param to pointer to the output buffer. The result is always |
181 | zero-terminated, and the longest returned string is thus |
182 | 'width + 1' bytes. |
183 | @param error if not NULL, points to a location where the status of |
184 | conversion is stored upon return. |
185 | FALSE successful conversion |
186 | TRUE the input number is [-,+]infinity or nan. |
187 | The output string in this case is always '0'. |
188 | @return number of written characters (excluding terminating '\0') |
189 | |
190 | @todo |
191 | Check if it is possible and makes sense to do our own rounding on top of |
192 | dtoa() instead of calling dtoa() twice in (rare) cases when the resulting |
193 | string representation does not fit in the specified field width and we want |
194 | to re-round the input number with fewer significant digits. Examples: |
195 | |
196 | my_gcvt(-9e-3, ..., 4, ...); |
197 | my_gcvt(-9e-3, ..., 2, ...); |
198 | my_gcvt(1.87e-3, ..., 4, ...); |
199 | my_gcvt(55, ..., 1, ...); |
200 | |
201 | We do our best to minimize such cases by: |
202 | |
203 | - passing to dtoa() the field width as the number of significant digits |
204 | |
205 | - removing the sign of the number early (and decreasing the width before |
206 | passing it to dtoa()) |
207 | |
208 | - choosing the proper format to preserve the most number of significant |
209 | digits. |
210 | */ |
211 | |
212 | size_t my_gcvt(double x, my_gcvt_arg_type type, int width, char *to, |
213 | my_bool *error) |
214 | { |
215 | int decpt, sign, len, exp_len; |
216 | char *res, *src, *end, *dst= to, *dend= dst + width; |
217 | char buf[DTOA_BUFF_SIZE]; |
218 | my_bool have_space, force_e_format; |
219 | DBUG_ASSERT(width > 0 && to != NULL); |
220 | |
221 | /* We want to remove '-' from equations early */ |
222 | if (x < 0.) |
223 | width--; |
224 | |
225 | res= dtoa(x, 4, type == MY_GCVT_ARG_DOUBLE ? width : MY_MIN(width, FLT_DIG), |
226 | &decpt, &sign, &end, buf, sizeof(buf)); |
227 | if (decpt == DTOA_OVERFLOW) |
228 | { |
229 | dtoa_free(res, buf, sizeof(buf)); |
230 | *to++= '0'; |
231 | *to= '\0'; |
232 | if (error != NULL) |
233 | *error= TRUE; |
234 | return 1; |
235 | } |
236 | |
237 | if (error != NULL) |
238 | *error= FALSE; |
239 | |
240 | src= res; |
241 | len= (int)(end - res); |
242 | |
243 | /* |
244 | Number of digits in the exponent from the 'e' conversion. |
245 | The sign of the exponent is taken into account separetely, we don't need |
246 | to count it here. |
247 | */ |
248 | exp_len= 1 + (decpt >= 101 || decpt <= -99) + (decpt >= 11 || decpt <= -9); |
249 | |
250 | /* |
251 | Do we have enough space for all digits in the 'f' format? |
252 | Let 'len' be the number of significant digits returned by dtoa, |
253 | and F be the length of the resulting decimal representation. |
254 | Consider the following cases: |
255 | 1. decpt <= 0, i.e. we have "0.NNN" => F = len - decpt + 2 |
256 | 2. 0 < decpt < len, i.e. we have "NNN.NNN" => F = len + 1 |
257 | 3. len <= decpt, i.e. we have "NNN00" => F = decpt |
258 | */ |
259 | have_space= (decpt <= 0 ? len - decpt + 2 : |
260 | decpt > 0 && decpt < len ? len + 1 : |
261 | decpt) <= width; |
262 | /* |
263 | The following is true when no significant digits can be placed with the |
264 | specified field width using the 'f' format, and the 'e' format |
265 | will not be truncated. |
266 | */ |
267 | force_e_format= (decpt <= 0 && width <= 2 - decpt && width >= 3 + exp_len); |
268 | /* |
269 | Assume that we don't have enough space to place all significant digits in |
270 | the 'f' format. We have to choose between the 'e' format and the 'f' one |
271 | to keep as many significant digits as possible. |
272 | Let E and F be the lengths of decimal representation in the 'e' and 'f' |
273 | formats, respectively. We want to use the 'f' format if, and only if F <= E. |
274 | Consider the following cases: |
275 | 1. decpt <= 0. |
276 | F = len - decpt + 2 (see above) |
277 | E = len + (len > 1) + 1 + 1 (decpt <= -99) + (decpt <= -9) + 1 |
278 | ("N.NNe-MMM") |
279 | (F <= E) <=> (len == 1 && decpt >= -1) || (len > 1 && decpt >= -2) |
280 | We also need to ensure that if the 'f' format is chosen, |
281 | the field width allows us to place at least one significant digit |
282 | (i.e. width > 2 - decpt). If not, we prefer the 'e' format. |
283 | 2. 0 < decpt < len |
284 | F = len + 1 (see above) |
285 | E = len + 1 + 1 + ... ("N.NNeMMM") |
286 | F is always less than E. |
287 | 3. len <= decpt <= width |
288 | In this case we have enough space to represent the number in the 'f' |
289 | format, so we prefer it with some exceptions. |
290 | 4. width < decpt |
291 | The number cannot be represented in the 'f' format at all, always use |
292 | the 'e' 'one. |
293 | */ |
294 | if ((have_space || |
295 | /* |
296 | Not enough space, let's see if the 'f' format provides the most number |
297 | of significant digits. |
298 | */ |
299 | ((decpt <= width && (decpt >= -1 || (decpt == -2 && |
300 | (len > 1 || !force_e_format)))) && |
301 | !force_e_format)) && |
302 | |
303 | /* |
304 | Use the 'e' format in some cases even if we have enough space for the |
305 | 'f' one. See comment for MAX_DECPT_FOR_F_FORMAT. |
306 | */ |
307 | (!have_space || (decpt >= -MAX_DECPT_FOR_F_FORMAT + 1 && |
308 | (decpt <= MAX_DECPT_FOR_F_FORMAT || len > decpt)))) |
309 | { |
310 | /* 'f' format */ |
311 | int i; |
312 | |
313 | width-= (decpt < len) + (decpt <= 0 ? 1 - decpt : 0); |
314 | |
315 | /* Do we have to truncate any digits? */ |
316 | if (width < len) |
317 | { |
318 | if (width < decpt) |
319 | { |
320 | if (error != NULL) |
321 | *error= TRUE; |
322 | width= decpt; |
323 | } |
324 | |
325 | /* |
326 | We want to truncate (len - width) least significant digits after the |
327 | decimal point. For this we are calling dtoa with mode=5, passing the |
328 | number of significant digits = (len-decpt) - (len-width) = width-decpt |
329 | */ |
330 | dtoa_free(res, buf, sizeof(buf)); |
331 | res= dtoa(x, 5, width - decpt, &decpt, &sign, &end, buf, sizeof(buf)); |
332 | src= res; |
333 | len= (int)(end - res); |
334 | } |
335 | |
336 | if (len == 0) |
337 | { |
338 | /* Underflow. Just print '0' and exit */ |
339 | *dst++= '0'; |
340 | goto end; |
341 | } |
342 | |
343 | /* |
344 | At this point we are sure we have enough space to put all digits |
345 | returned by dtoa |
346 | */ |
347 | if (sign && dst < dend) |
348 | *dst++= '-'; |
349 | if (decpt <= 0) |
350 | { |
351 | if (dst < dend) |
352 | *dst++= '0'; |
353 | if (len > 0 && dst < dend) |
354 | *dst++= '.'; |
355 | for (; decpt < 0 && dst < dend; decpt++) |
356 | *dst++= '0'; |
357 | } |
358 | |
359 | for (i= 1; i <= len && dst < dend; i++) |
360 | { |
361 | *dst++= *src++; |
362 | if (i == decpt && i < len && dst < dend) |
363 | *dst++= '.'; |
364 | } |
365 | while (i++ <= decpt && dst < dend) |
366 | *dst++= '0'; |
367 | } |
368 | else |
369 | { |
370 | /* 'e' format */ |
371 | int decpt_sign= 0; |
372 | |
373 | if (--decpt < 0) |
374 | { |
375 | decpt= -decpt; |
376 | width--; |
377 | decpt_sign= 1; |
378 | } |
379 | width-= 1 + exp_len; /* eNNN */ |
380 | |
381 | if (len > 1) |
382 | width--; |
383 | |
384 | if (width <= 0) |
385 | { |
386 | /* Overflow */ |
387 | if (error != NULL) |
388 | *error= TRUE; |
389 | width= 0; |
390 | } |
391 | |
392 | /* Do we have to truncate any digits? */ |
393 | if (width < len) |
394 | { |
395 | /* Yes, re-convert with a smaller width */ |
396 | dtoa_free(res, buf, sizeof(buf)); |
397 | res= dtoa(x, 4, width, &decpt, &sign, &end, buf, sizeof(buf)); |
398 | src= res; |
399 | len= (int)(end - res); |
400 | if (--decpt < 0) |
401 | decpt= -decpt; |
402 | } |
403 | /* |
404 | At this point we are sure we have enough space to put all digits |
405 | returned by dtoa |
406 | */ |
407 | if (sign && dst < dend) |
408 | *dst++= '-'; |
409 | if (dst < dend) |
410 | *dst++= *src++; |
411 | if (len > 1 && dst < dend) |
412 | { |
413 | *dst++= '.'; |
414 | while (src < end && dst < dend) |
415 | *dst++= *src++; |
416 | } |
417 | if (dst < dend) |
418 | *dst++= 'e'; |
419 | if (decpt_sign && dst < dend) |
420 | *dst++= '-'; |
421 | |
422 | if (decpt >= 100 && dst < dend) |
423 | { |
424 | *dst++= decpt / 100 + '0'; |
425 | decpt%= 100; |
426 | if (dst < dend) |
427 | *dst++= decpt / 10 + '0'; |
428 | } |
429 | else if (decpt >= 10 && dst < dend) |
430 | *dst++= decpt / 10 + '0'; |
431 | if (dst < dend) |
432 | *dst++= decpt % 10 + '0'; |
433 | |
434 | } |
435 | |
436 | end: |
437 | dtoa_free(res, buf, sizeof(buf)); |
438 | *dst= '\0'; |
439 | |
440 | return dst - to; |
441 | } |
442 | |
443 | /** |
444 | @brief |
445 | Converts string to double (string does not have to be zero-terminated) |
446 | |
447 | @details |
448 | This is a wrapper around dtoa's version of strtod(). |
449 | |
450 | @param str input string |
451 | @param end address of a pointer to the first character after the input |
452 | string. Upon return the pointer is set to point to the first |
453 | rejected character. |
454 | @param error Upon return is set to EOVERFLOW in case of underflow or |
455 | overflow. |
456 | |
457 | @return The resulting double value. In case of underflow, 0.0 is |
458 | returned. In case overflow, signed DBL_MAX is returned. |
459 | */ |
460 | |
461 | double my_strtod(const char *str, char **end, int *error) |
462 | { |
463 | char buf[DTOA_BUFF_SIZE]; |
464 | double res; |
465 | DBUG_ASSERT(end != NULL && ((str != NULL && *end != NULL) || |
466 | (str == NULL && *end == NULL)) && |
467 | error != NULL); |
468 | |
469 | res= my_strtod_int(str, end, error, buf, sizeof(buf)); |
470 | return (*error == 0) ? res : (res < 0 ? -DBL_MAX : DBL_MAX); |
471 | } |
472 | |
473 | |
474 | double my_atof(const char *nptr) |
475 | { |
476 | int error; |
477 | const char *end= nptr+65535; /* Should be enough */ |
478 | return (my_strtod(nptr, (char**) &end, &error)); |
479 | } |
480 | |
481 | |
482 | /**************************************************************** |
483 | * |
484 | * The author of this software is David M. Gay. |
485 | * |
486 | * Copyright (c) 1991, 2000, 2001 by Lucent Technologies. |
487 | * |
488 | * Permission to use, copy, modify, and distribute this software for any |
489 | * purpose without fee is hereby granted, provided that this entire notice |
490 | * is included in all copies of any software which is or includes a copy |
491 | * or modification of this software and in all copies of the supporting |
492 | * documentation for such software. |
493 | * |
494 | * THIS SOFTWARE IS BEING PROVIDED "AS IS", WITHOUT ANY EXPRESS OR IMPLIED |
495 | * WARRANTY. IN PARTICULAR, NEITHER THE AUTHOR NOR LUCENT MAKES ANY |
496 | * REPRESENTATION OR WARRANTY OF ANY KIND CONCERNING THE MERCHANTABILITY |
497 | * OF THIS SOFTWARE OR ITS FITNESS FOR ANY PARTICULAR PURPOSE. |
498 | * |
499 | ***************************************************************/ |
500 | /* Please send bug reports to David M. Gay (dmg at acm dot org, |
501 | * with " at " changed at "@" and " dot " changed to "."). */ |
502 | |
503 | /* |
504 | Original copy of the software is located at http://www.netlib.org/fp/dtoa.c |
505 | It was adjusted to serve MySQL server needs: |
506 | * strtod() was modified to not expect a zero-terminated string. |
507 | It now honors 'se' (end of string) argument as the input parameter, |
508 | not just as the output one. |
509 | * in dtoa(), in case of overflow/underflow/NaN result string now contains "0"; |
510 | decpt is set to DTOA_OVERFLOW to indicate overflow. |
511 | * support for VAX, IBM mainframe and 16-bit hardware removed |
512 | * we always assume that 64-bit integer type is available |
513 | * support for Kernigan-Ritchie style headers (pre-ANSI compilers) |
514 | removed |
515 | * all gcc warnings ironed out |
516 | * we always assume multithreaded environment, so we had to change |
517 | memory allocation procedures to use stack in most cases; |
518 | malloc is used as the last resort. |
519 | * pow5mult rewritten to use pre-calculated pow5 list instead of |
520 | the one generated on the fly. |
521 | */ |
522 | |
523 | |
524 | /* |
525 | On a machine with IEEE extended-precision registers, it is |
526 | necessary to specify double-precision (53-bit) rounding precision |
527 | before invoking strtod or dtoa. If the machine uses (the equivalent |
528 | of) Intel 80x87 arithmetic, the call |
529 | _control87(PC_53, MCW_PC); |
530 | does this with many compilers. Whether this or another call is |
531 | appropriate depends on the compiler; for this to work, it may be |
532 | necessary to #include "float.h" or another system-dependent header |
533 | file. |
534 | */ |
535 | |
536 | /* |
537 | #define Honor_FLT_ROUNDS if FLT_ROUNDS can assume the values 2 or 3 |
538 | and dtoa should round accordingly. |
539 | #define Check_FLT_ROUNDS if FLT_ROUNDS can assume the values 2 or 3 |
540 | and Honor_FLT_ROUNDS is not #defined. |
541 | |
542 | TODO: check if we can get rid of the above two |
543 | */ |
544 | |
545 | typedef int32 Long; |
546 | typedef uint32 ULong; |
547 | typedef int64 LLong; |
548 | typedef uint64 ULLong; |
549 | |
550 | typedef union { double d; ULong L[2]; } U; |
551 | |
552 | #if defined(WORDS_BIGENDIAN) || (defined(__FLOAT_WORD_ORDER) && \ |
553 | (__FLOAT_WORD_ORDER == __BIG_ENDIAN)) |
554 | #define word0(x) (x)->L[0] |
555 | #define word1(x) (x)->L[1] |
556 | #else |
557 | #define word0(x) (x)->L[1] |
558 | #define word1(x) (x)->L[0] |
559 | #endif |
560 | |
561 | #define dval(x) (x)->d |
562 | |
563 | /* #define P DBL_MANT_DIG */ |
564 | /* Ten_pmax= floor(P*log(2)/log(5)) */ |
565 | /* Bletch= (highest power of 2 < DBL_MAX_10_EXP) / 16 */ |
566 | /* Quick_max= floor((P-1)*log(FLT_RADIX)/log(10) - 1) */ |
567 | /* Int_max= floor(P*log(FLT_RADIX)/log(10) - 1) */ |
568 | |
569 | #define Exp_shift 20 |
570 | #define Exp_shift1 20 |
571 | #define Exp_msk1 0x100000 |
572 | #define Exp_mask 0x7ff00000 |
573 | #define P 53 |
574 | #define Bias 1023 |
575 | #define Emin (-1022) |
576 | #define Exp_1 0x3ff00000 |
577 | #define Exp_11 0x3ff00000 |
578 | #define Ebits 11 |
579 | #define Frac_mask 0xfffff |
580 | #define Frac_mask1 0xfffff |
581 | #define Ten_pmax 22 |
582 | #define Bletch 0x10 |
583 | #define Bndry_mask 0xfffff |
584 | #define Bndry_mask1 0xfffff |
585 | #define LSB 1 |
586 | #define Sign_bit 0x80000000 |
587 | #define Log2P 1 |
588 | #define Tiny1 1 |
589 | #define Quick_max 14 |
590 | #define Int_max 14 |
591 | |
592 | #ifndef Flt_Rounds |
593 | #ifdef FLT_ROUNDS |
594 | #define Flt_Rounds FLT_ROUNDS |
595 | #else |
596 | #define Flt_Rounds 1 |
597 | #endif |
598 | #endif /*Flt_Rounds*/ |
599 | |
600 | #ifdef Honor_FLT_ROUNDS |
601 | #define Rounding rounding |
602 | #undef Check_FLT_ROUNDS |
603 | #define Check_FLT_ROUNDS |
604 | #else |
605 | #define Rounding Flt_Rounds |
606 | #endif |
607 | |
608 | #define rounded_product(a,b) a*= b |
609 | #define rounded_quotient(a,b) a/= b |
610 | |
611 | #define Big0 (Frac_mask1 | Exp_msk1*(DBL_MAX_EXP+Bias-1)) |
612 | #define Big1 0xffffffff |
613 | #define FFFFFFFF 0xffffffffUL |
614 | |
615 | /* This is tested to be enough for dtoa */ |
616 | |
617 | #define Kmax 15 |
618 | |
619 | #define Bcopy(x,y) memcpy((char *)&x->sign, (char *)&y->sign, \ |
620 | 2*sizeof(int) + y->wds*sizeof(ULong)) |
621 | |
622 | /* Arbitrary-length integer */ |
623 | |
624 | typedef struct Bigint |
625 | { |
626 | union { |
627 | ULong *x; /* points right after this Bigint object */ |
628 | struct Bigint *next; /* to maintain free lists */ |
629 | } p; |
630 | int k; /* 2^k = maxwds */ |
631 | int maxwds; /* maximum length in 32-bit words */ |
632 | int sign; /* not zero if number is negative */ |
633 | int wds; /* current length in 32-bit words */ |
634 | } Bigint; |
635 | |
636 | |
637 | /* A simple stack-memory based allocator for Bigints */ |
638 | |
639 | typedef struct Stack_alloc |
640 | { |
641 | char *begin; |
642 | char *free; |
643 | char *end; |
644 | /* |
645 | Having list of free blocks lets us reduce maximum required amount |
646 | of memory from ~4000 bytes to < 1680 (tested on x86). |
647 | */ |
648 | Bigint *freelist[Kmax+1]; |
649 | } Stack_alloc; |
650 | |
651 | |
652 | /* |
653 | Try to allocate object on stack, and resort to malloc if all |
654 | stack memory is used. Ensure allocated objects to be aligned by the pointer |
655 | size in order to not break the alignment rules when storing a pointer to a |
656 | Bigint. |
657 | */ |
658 | |
659 | static Bigint *Balloc(int k, Stack_alloc *alloc) |
660 | { |
661 | Bigint *rv; |
662 | DBUG_ASSERT(k <= Kmax); |
663 | if (k <= Kmax && alloc->freelist[k]) |
664 | { |
665 | rv= alloc->freelist[k]; |
666 | alloc->freelist[k]= rv->p.next; |
667 | } |
668 | else |
669 | { |
670 | int x, len; |
671 | |
672 | x= 1 << k; |
673 | len= MY_ALIGN(sizeof(Bigint) + x * sizeof(ULong), SIZEOF_CHARP); |
674 | |
675 | if (alloc->free + len <= alloc->end) |
676 | { |
677 | rv= (Bigint*) alloc->free; |
678 | alloc->free+= len; |
679 | } |
680 | else |
681 | rv= (Bigint*) malloc(len); |
682 | |
683 | rv->k= k; |
684 | rv->maxwds= x; |
685 | } |
686 | rv->sign= rv->wds= 0; |
687 | rv->p.x= (ULong*) (rv + 1); |
688 | return rv; |
689 | } |
690 | |
691 | |
692 | /* |
693 | If object was allocated on stack, try putting it to the free |
694 | list. Otherwise call free(). |
695 | */ |
696 | |
697 | static void Bfree(Bigint *v, Stack_alloc *alloc) |
698 | { |
699 | char *gptr= (char*) v; /* generic pointer */ |
700 | if (gptr < alloc->begin || gptr >= alloc->end) |
701 | free(gptr); |
702 | else if (v->k <= Kmax) |
703 | { |
704 | /* |
705 | Maintain free lists only for stack objects: this way we don't |
706 | have to bother with freeing lists in the end of dtoa; |
707 | heap should not be used normally anyway. |
708 | */ |
709 | v->p.next= alloc->freelist[v->k]; |
710 | alloc->freelist[v->k]= v; |
711 | } |
712 | } |
713 | |
714 | |
715 | /* |
716 | This is to place return value of dtoa in: tries to use stack |
717 | as well, but passes by free lists management and just aligns len by |
718 | the pointer size in order to not break the alignment rules when storing a |
719 | pointer to a Bigint. |
720 | */ |
721 | |
722 | static char *dtoa_alloc(int i, Stack_alloc *alloc) |
723 | { |
724 | char *rv; |
725 | int aligned_size= MY_ALIGN(i, SIZEOF_CHARP); |
726 | if (alloc->free + aligned_size <= alloc->end) |
727 | { |
728 | rv= alloc->free; |
729 | alloc->free+= aligned_size; |
730 | } |
731 | else |
732 | rv= malloc(i); |
733 | return rv; |
734 | } |
735 | |
736 | |
737 | /* |
738 | dtoa_free() must be used to free values s returned by dtoa() |
739 | This is the counterpart of dtoa_alloc() |
740 | */ |
741 | |
742 | static void dtoa_free(char *gptr, char *buf, size_t buf_size) |
743 | { |
744 | if (gptr < buf || gptr >= buf + buf_size) |
745 | free(gptr); |
746 | } |
747 | |
748 | |
749 | /* Bigint arithmetic functions */ |
750 | |
751 | /* Multiply by m and add a */ |
752 | |
753 | static Bigint *multadd(Bigint *b, int m, int a, Stack_alloc *alloc) |
754 | { |
755 | int i, wds; |
756 | ULong *x; |
757 | ULLong carry, y; |
758 | Bigint *b1; |
759 | |
760 | wds= b->wds; |
761 | x= b->p.x; |
762 | i= 0; |
763 | carry= a; |
764 | do |
765 | { |
766 | y= *x * (ULLong)m + carry; |
767 | carry= y >> 32; |
768 | *x++= (ULong)(y & FFFFFFFF); |
769 | } |
770 | while (++i < wds); |
771 | if (carry) |
772 | { |
773 | if (wds >= b->maxwds) |
774 | { |
775 | b1= Balloc(b->k+1, alloc); |
776 | Bcopy(b1, b); |
777 | Bfree(b, alloc); |
778 | b= b1; |
779 | } |
780 | b->p.x[wds++]= (ULong) carry; |
781 | b->wds= wds; |
782 | } |
783 | return b; |
784 | } |
785 | |
786 | /** |
787 | Converts a string to Bigint. |
788 | |
789 | Now we have nd0 digits, starting at s, followed by a |
790 | decimal point, followed by nd-nd0 digits. |
791 | Unless nd0 == nd, in which case we have a number of the form: |
792 | ".xxxxxx" or "xxxxxx." |
793 | |
794 | @param s Input string, already partially parsed by my_strtod_int(). |
795 | @param nd0 Number of digits before decimal point. |
796 | @param nd Total number of digits. |
797 | @param y9 Pre-computed value of the first nine digits. |
798 | @param alloc Stack allocator for Bigints. |
799 | */ |
800 | static Bigint *s2b(const char *s, int nd0, int nd, ULong y9, Stack_alloc *alloc) |
801 | { |
802 | Bigint *b; |
803 | int i, k; |
804 | Long x, y; |
805 | |
806 | x= (nd + 8) / 9; |
807 | for (k= 0, y= 1; x > y; y <<= 1, k++) ; |
808 | b= Balloc(k, alloc); |
809 | b->p.x[0]= y9; |
810 | b->wds= 1; |
811 | |
812 | i= 9; |
813 | if (9 < nd0) |
814 | { |
815 | s+= 9; |
816 | do |
817 | b= multadd(b, 10, *s++ - '0', alloc); |
818 | while (++i < nd0); |
819 | s++; /* skip '.' */ |
820 | } |
821 | else |
822 | s+= 10; |
823 | /* now do the fractional part */ |
824 | for(; i < nd; i++) |
825 | b= multadd(b, 10, *s++ - '0', alloc); |
826 | return b; |
827 | } |
828 | |
829 | |
830 | static int hi0bits(register ULong x) |
831 | { |
832 | register int k= 0; |
833 | |
834 | if (!(x & 0xffff0000)) |
835 | { |
836 | k= 16; |
837 | x<<= 16; |
838 | } |
839 | if (!(x & 0xff000000)) |
840 | { |
841 | k+= 8; |
842 | x<<= 8; |
843 | } |
844 | if (!(x & 0xf0000000)) |
845 | { |
846 | k+= 4; |
847 | x<<= 4; |
848 | } |
849 | if (!(x & 0xc0000000)) |
850 | { |
851 | k+= 2; |
852 | x<<= 2; |
853 | } |
854 | if (!(x & 0x80000000)) |
855 | { |
856 | k++; |
857 | if (!(x & 0x40000000)) |
858 | return 32; |
859 | } |
860 | return k; |
861 | } |
862 | |
863 | |
864 | static int lo0bits(ULong *y) |
865 | { |
866 | register int k; |
867 | register ULong x= *y; |
868 | |
869 | if (x & 7) |
870 | { |
871 | if (x & 1) |
872 | return 0; |
873 | if (x & 2) |
874 | { |
875 | *y= x >> 1; |
876 | return 1; |
877 | } |
878 | *y= x >> 2; |
879 | return 2; |
880 | } |
881 | k= 0; |
882 | if (!(x & 0xffff)) |
883 | { |
884 | k= 16; |
885 | x>>= 16; |
886 | } |
887 | if (!(x & 0xff)) |
888 | { |
889 | k+= 8; |
890 | x>>= 8; |
891 | } |
892 | if (!(x & 0xf)) |
893 | { |
894 | k+= 4; |
895 | x>>= 4; |
896 | } |
897 | if (!(x & 0x3)) |
898 | { |
899 | k+= 2; |
900 | x>>= 2; |
901 | } |
902 | if (!(x & 1)) |
903 | { |
904 | k++; |
905 | x>>= 1; |
906 | if (!x) |
907 | return 32; |
908 | } |
909 | *y= x; |
910 | return k; |
911 | } |
912 | |
913 | |
914 | /* Convert integer to Bigint number */ |
915 | |
916 | static Bigint *i2b(int i, Stack_alloc *alloc) |
917 | { |
918 | Bigint *b; |
919 | |
920 | b= Balloc(1, alloc); |
921 | b->p.x[0]= i; |
922 | b->wds= 1; |
923 | return b; |
924 | } |
925 | |
926 | |
927 | /* Multiply two Bigint numbers */ |
928 | |
929 | static Bigint *mult(Bigint *a, Bigint *b, Stack_alloc *alloc) |
930 | { |
931 | Bigint *c; |
932 | int k, wa, wb, wc; |
933 | ULong *x, *xa, *xae, *xb, *xbe, *xc, *xc0; |
934 | ULong y; |
935 | ULLong carry, z; |
936 | |
937 | if (a->wds < b->wds) |
938 | { |
939 | c= a; |
940 | a= b; |
941 | b= c; |
942 | } |
943 | k= a->k; |
944 | wa= a->wds; |
945 | wb= b->wds; |
946 | wc= wa + wb; |
947 | if (wc > a->maxwds) |
948 | k++; |
949 | c= Balloc(k, alloc); |
950 | for (x= c->p.x, xa= x + wc; x < xa; x++) |
951 | *x= 0; |
952 | xa= a->p.x; |
953 | xae= xa + wa; |
954 | xb= b->p.x; |
955 | xbe= xb + wb; |
956 | xc0= c->p.x; |
957 | for (; xb < xbe; xc0++) |
958 | { |
959 | if ((y= *xb++)) |
960 | { |
961 | x= xa; |
962 | xc= xc0; |
963 | carry= 0; |
964 | do |
965 | { |
966 | z= *x++ * (ULLong)y + *xc + carry; |
967 | carry= z >> 32; |
968 | *xc++= (ULong) (z & FFFFFFFF); |
969 | } |
970 | while (x < xae); |
971 | *xc= (ULong) carry; |
972 | } |
973 | } |
974 | for (xc0= c->p.x, xc= xc0 + wc; wc > 0 && !*--xc; --wc) ; |
975 | c->wds= wc; |
976 | return c; |
977 | } |
978 | |
979 | |
980 | /* |
981 | Precalculated array of powers of 5: tested to be enough for |
982 | vasting majority of dtoa_r cases. |
983 | */ |
984 | |
985 | static ULong powers5[]= |
986 | { |
987 | 625UL, |
988 | |
989 | 390625UL, |
990 | |
991 | 2264035265UL, 35UL, |
992 | |
993 | 2242703233UL, 762134875UL, 1262UL, |
994 | |
995 | 3211403009UL, 1849224548UL, 3668416493UL, 3913284084UL, 1593091UL, |
996 | |
997 | 781532673UL, 64985353UL, 253049085UL, 594863151UL, 3553621484UL, |
998 | 3288652808UL, 3167596762UL, 2788392729UL, 3911132675UL, 590UL, |
999 | |
1000 | 2553183233UL, 3201533787UL, 3638140786UL, 303378311UL, 1809731782UL, |
1001 | 3477761648UL, 3583367183UL, 649228654UL, 2915460784UL, 487929380UL, |
1002 | 1011012442UL, 1677677582UL, 3428152256UL, 1710878487UL, 1438394610UL, |
1003 | 2161952759UL, 4100910556UL, 1608314830UL, 349175UL |
1004 | }; |
1005 | |
1006 | |
1007 | static Bigint p5_a[]= |
1008 | { |
1009 | /* { x } - k - maxwds - sign - wds */ |
1010 | { { powers5 }, 1, 1, 0, 1 }, |
1011 | { { powers5 + 1 }, 1, 1, 0, 1 }, |
1012 | { { powers5 + 2 }, 1, 2, 0, 2 }, |
1013 | { { powers5 + 4 }, 2, 3, 0, 3 }, |
1014 | { { powers5 + 7 }, 3, 5, 0, 5 }, |
1015 | { { powers5 + 12 }, 4, 10, 0, 10 }, |
1016 | { { powers5 + 22 }, 5, 19, 0, 19 } |
1017 | }; |
1018 | |
1019 | #define P5A_MAX (sizeof(p5_a)/sizeof(*p5_a) - 1) |
1020 | |
1021 | static Bigint *pow5mult(Bigint *b, int k, Stack_alloc *alloc) |
1022 | { |
1023 | Bigint *b1, *p5, *p51=NULL; |
1024 | int i; |
1025 | static int p05[3]= { 5, 25, 125 }; |
1026 | my_bool overflow= FALSE; |
1027 | |
1028 | if ((i= k & 3)) |
1029 | b= multadd(b, p05[i-1], 0, alloc); |
1030 | |
1031 | if (!(k>>= 2)) |
1032 | return b; |
1033 | p5= p5_a; |
1034 | for (;;) |
1035 | { |
1036 | if (k & 1) |
1037 | { |
1038 | b1= mult(b, p5, alloc); |
1039 | Bfree(b, alloc); |
1040 | b= b1; |
1041 | } |
1042 | if (!(k>>= 1)) |
1043 | break; |
1044 | /* Calculate next power of 5 */ |
1045 | if (overflow) |
1046 | { |
1047 | p51= mult(p5, p5, alloc); |
1048 | Bfree(p5, alloc); |
1049 | p5= p51; |
1050 | } |
1051 | else if (p5 < p5_a + P5A_MAX) |
1052 | ++p5; |
1053 | else if (p5 == p5_a + P5A_MAX) |
1054 | { |
1055 | p5= mult(p5, p5, alloc); |
1056 | overflow= TRUE; |
1057 | } |
1058 | } |
1059 | if (p51) |
1060 | Bfree(p51, alloc); |
1061 | return b; |
1062 | } |
1063 | |
1064 | |
1065 | static Bigint *lshift(Bigint *b, int k, Stack_alloc *alloc) |
1066 | { |
1067 | int i, k1, n, n1; |
1068 | Bigint *b1; |
1069 | ULong *x, *x1, *xe, z; |
1070 | |
1071 | n= k >> 5; |
1072 | k1= b->k; |
1073 | n1= n + b->wds + 1; |
1074 | for (i= b->maxwds; n1 > i; i<<= 1) |
1075 | k1++; |
1076 | b1= Balloc(k1, alloc); |
1077 | x1= b1->p.x; |
1078 | for (i= 0; i < n; i++) |
1079 | *x1++= 0; |
1080 | x= b->p.x; |
1081 | xe= x + b->wds; |
1082 | if (k&= 0x1f) |
1083 | { |
1084 | k1= 32 - k; |
1085 | z= 0; |
1086 | do |
1087 | { |
1088 | *x1++= *x << k | z; |
1089 | z= *x++ >> k1; |
1090 | } |
1091 | while (x < xe); |
1092 | if ((*x1= z)) |
1093 | ++n1; |
1094 | } |
1095 | else |
1096 | do |
1097 | *x1++= *x++; |
1098 | while (x < xe); |
1099 | b1->wds= n1 - 1; |
1100 | Bfree(b, alloc); |
1101 | return b1; |
1102 | } |
1103 | |
1104 | |
1105 | static int cmp(Bigint *a, Bigint *b) |
1106 | { |
1107 | ULong *xa, *xa0, *xb, *xb0; |
1108 | int i, j; |
1109 | |
1110 | i= a->wds; |
1111 | j= b->wds; |
1112 | if (i-= j) |
1113 | return i; |
1114 | xa0= a->p.x; |
1115 | xa= xa0 + j; |
1116 | xb0= b->p.x; |
1117 | xb= xb0 + j; |
1118 | for (;;) |
1119 | { |
1120 | if (*--xa != *--xb) |
1121 | return *xa < *xb ? -1 : 1; |
1122 | if (xa <= xa0) |
1123 | break; |
1124 | } |
1125 | return 0; |
1126 | } |
1127 | |
1128 | |
1129 | static Bigint *diff(Bigint *a, Bigint *b, Stack_alloc *alloc) |
1130 | { |
1131 | Bigint *c; |
1132 | int i, wa, wb; |
1133 | ULong *xa, *xae, *xb, *xbe, *xc; |
1134 | ULLong borrow, y; |
1135 | |
1136 | i= cmp(a,b); |
1137 | if (!i) |
1138 | { |
1139 | c= Balloc(0, alloc); |
1140 | c->wds= 1; |
1141 | c->p.x[0]= 0; |
1142 | return c; |
1143 | } |
1144 | if (i < 0) |
1145 | { |
1146 | c= a; |
1147 | a= b; |
1148 | b= c; |
1149 | i= 1; |
1150 | } |
1151 | else |
1152 | i= 0; |
1153 | c= Balloc(a->k, alloc); |
1154 | c->sign= i; |
1155 | wa= a->wds; |
1156 | xa= a->p.x; |
1157 | xae= xa + wa; |
1158 | wb= b->wds; |
1159 | xb= b->p.x; |
1160 | xbe= xb + wb; |
1161 | xc= c->p.x; |
1162 | borrow= 0; |
1163 | do |
1164 | { |
1165 | y= (ULLong)*xa++ - *xb++ - borrow; |
1166 | borrow= y >> 32 & (ULong)1; |
1167 | *xc++= (ULong) (y & FFFFFFFF); |
1168 | } |
1169 | while (xb < xbe); |
1170 | while (xa < xae) |
1171 | { |
1172 | y= *xa++ - borrow; |
1173 | borrow= y >> 32 & (ULong)1; |
1174 | *xc++= (ULong) (y & FFFFFFFF); |
1175 | } |
1176 | while (!*--xc) |
1177 | wa--; |
1178 | c->wds= wa; |
1179 | return c; |
1180 | } |
1181 | |
1182 | |
1183 | static double ulp(U *x) |
1184 | { |
1185 | register Long L; |
1186 | U u; |
1187 | |
1188 | L= (word0(x) & Exp_mask) - (P - 1)*Exp_msk1; |
1189 | word0(&u) = L; |
1190 | word1(&u) = 0; |
1191 | return dval(&u); |
1192 | } |
1193 | |
1194 | |
1195 | static double b2d(Bigint *a, int *e) |
1196 | { |
1197 | ULong *xa, *xa0, w, y, z; |
1198 | int k; |
1199 | U d; |
1200 | #define d0 word0(&d) |
1201 | #define d1 word1(&d) |
1202 | |
1203 | xa0= a->p.x; |
1204 | xa= xa0 + a->wds; |
1205 | y= *--xa; |
1206 | k= hi0bits(y); |
1207 | *e= 32 - k; |
1208 | if (k < Ebits) |
1209 | { |
1210 | d0= Exp_1 | y >> (Ebits - k); |
1211 | w= xa > xa0 ? *--xa : 0; |
1212 | d1= y << ((32-Ebits) + k) | w >> (Ebits - k); |
1213 | goto ret_d; |
1214 | } |
1215 | z= xa > xa0 ? *--xa : 0; |
1216 | if (k-= Ebits) |
1217 | { |
1218 | d0= Exp_1 | y << k | z >> (32 - k); |
1219 | y= xa > xa0 ? *--xa : 0; |
1220 | d1= z << k | y >> (32 - k); |
1221 | } |
1222 | else |
1223 | { |
1224 | d0= Exp_1 | y; |
1225 | d1= z; |
1226 | } |
1227 | ret_d: |
1228 | #undef d0 |
1229 | #undef d1 |
1230 | return dval(&d); |
1231 | } |
1232 | |
1233 | |
1234 | static Bigint *d2b(U *d, int *e, int *bits, Stack_alloc *alloc) |
1235 | { |
1236 | Bigint *b; |
1237 | int de, k; |
1238 | ULong *x, y, z; |
1239 | int i; |
1240 | #define d0 word0(d) |
1241 | #define d1 word1(d) |
1242 | |
1243 | b= Balloc(1, alloc); |
1244 | x= b->p.x; |
1245 | |
1246 | z= d0 & Frac_mask; |
1247 | d0 &= 0x7fffffff; /* clear sign bit, which we ignore */ |
1248 | if ((de= (int)(d0 >> Exp_shift))) |
1249 | z|= Exp_msk1; |
1250 | if ((y= d1)) |
1251 | { |
1252 | if ((k= lo0bits(&y))) |
1253 | { |
1254 | x[0]= y | z << (32 - k); |
1255 | z>>= k; |
1256 | } |
1257 | else |
1258 | x[0]= y; |
1259 | i= b->wds= (x[1]= z) ? 2 : 1; |
1260 | } |
1261 | else |
1262 | { |
1263 | k= lo0bits(&z); |
1264 | x[0]= z; |
1265 | i= b->wds= 1; |
1266 | k+= 32; |
1267 | } |
1268 | if (de) |
1269 | { |
1270 | *e= de - Bias - (P-1) + k; |
1271 | *bits= P - k; |
1272 | } |
1273 | else |
1274 | { |
1275 | *e= de - Bias - (P-1) + 1 + k; |
1276 | *bits= 32*i - hi0bits(x[i-1]); |
1277 | } |
1278 | return b; |
1279 | #undef d0 |
1280 | #undef d1 |
1281 | } |
1282 | |
1283 | |
1284 | static double ratio(Bigint *a, Bigint *b) |
1285 | { |
1286 | U da, db; |
1287 | int k, ka, kb; |
1288 | |
1289 | dval(&da)= b2d(a, &ka); |
1290 | dval(&db)= b2d(b, &kb); |
1291 | k= ka - kb + 32*(a->wds - b->wds); |
1292 | if (k > 0) |
1293 | word0(&da)+= (ULong)(k*Exp_msk1 * 1.0); |
1294 | else |
1295 | { |
1296 | k= -k; |
1297 | word0(&db)+= k*Exp_msk1; |
1298 | } |
1299 | return dval(&da) / dval(&db); |
1300 | } |
1301 | |
1302 | static const double tens[] = |
1303 | { |
1304 | 1e0, 1e1, 1e2, 1e3, 1e4, 1e5, 1e6, 1e7, 1e8, 1e9, |
1305 | 1e10, 1e11, 1e12, 1e13, 1e14, 1e15, 1e16, 1e17, 1e18, 1e19, |
1306 | 1e20, 1e21, 1e22 |
1307 | }; |
1308 | |
1309 | static const double bigtens[]= { 1e16, 1e32, 1e64, 1e128, 1e256 }; |
1310 | static const double tinytens[]= |
1311 | { 1e-16, 1e-32, 1e-64, 1e-128, |
1312 | 9007199254740992.*9007199254740992.e-256 /* = 2^106 * 1e-53 */ |
1313 | }; |
1314 | /* |
1315 | The factor of 2^53 in tinytens[4] helps us avoid setting the underflow |
1316 | flag unnecessarily. It leads to a song and dance at the end of strtod. |
1317 | */ |
1318 | #define Scale_Bit 0x10 |
1319 | #define n_bigtens 5 |
1320 | |
1321 | /* |
1322 | strtod for IEEE--arithmetic machines. |
1323 | |
1324 | This strtod returns a nearest machine number to the input decimal |
1325 | string (or sets errno to EOVERFLOW). Ties are broken by the IEEE round-even |
1326 | rule. |
1327 | |
1328 | Inspired loosely by William D. Clinger's paper "How to Read Floating |
1329 | Point Numbers Accurately" [Proc. ACM SIGPLAN '90, pp. 92-101]. |
1330 | |
1331 | Modifications: |
1332 | |
1333 | 1. We only require IEEE (not IEEE double-extended). |
1334 | 2. We get by with floating-point arithmetic in a case that |
1335 | Clinger missed -- when we're computing d * 10^n |
1336 | for a small integer d and the integer n is not too |
1337 | much larger than 22 (the maximum integer k for which |
1338 | we can represent 10^k exactly), we may be able to |
1339 | compute (d*10^k) * 10^(e-k) with just one roundoff. |
1340 | 3. Rather than a bit-at-a-time adjustment of the binary |
1341 | result in the hard case, we use floating-point |
1342 | arithmetic to determine the adjustment to within |
1343 | one bit; only in really hard cases do we need to |
1344 | compute a second residual. |
1345 | 4. Because of 3., we don't need a large table of powers of 10 |
1346 | for ten-to-e (just some small tables, e.g. of 10^k |
1347 | for 0 <= k <= 22). |
1348 | */ |
1349 | |
1350 | static double my_strtod_int(const char *s00, char **se, int *error, char *buf, size_t buf_size) |
1351 | { |
1352 | int scale; |
1353 | int bb2, bb5, bbe, bd2, bd5, bbbits, bs2, UNINIT_VAR(c), dsign, |
1354 | e, e1, esign, i, j, k, nd, nd0, nf, nz, nz0, sign; |
1355 | const char *s, *s0, *s1, *end = *se; |
1356 | double aadj, aadj1; |
1357 | U aadj2, adj, rv, rv0; |
1358 | Long L; |
1359 | ULong y, z; |
1360 | Bigint *bb, *bb1, *bd, *bd0, *bs, *delta; |
1361 | #ifdef SET_INEXACT |
1362 | int inexact, oldinexact; |
1363 | #endif |
1364 | #ifdef Honor_FLT_ROUNDS |
1365 | int rounding; |
1366 | #endif |
1367 | Stack_alloc alloc; |
1368 | |
1369 | *error= 0; |
1370 | |
1371 | alloc.begin= alloc.free= buf; |
1372 | alloc.end= buf + buf_size; |
1373 | memset(alloc.freelist, 0, sizeof(alloc.freelist)); |
1374 | |
1375 | sign= nz0= nz= 0; |
1376 | dval(&rv)= 0.; |
1377 | for (s= s00; s < end; s++) |
1378 | switch (*s) { |
1379 | case '-': |
1380 | sign= 1; |
1381 | /* fall through */ |
1382 | case '+': |
1383 | s++; |
1384 | goto break2; |
1385 | case '\t': |
1386 | case '\n': |
1387 | case '\v': |
1388 | case '\f': |
1389 | case '\r': |
1390 | case ' ': |
1391 | continue; |
1392 | default: |
1393 | goto break2; |
1394 | } |
1395 | break2: |
1396 | if (s >= end) |
1397 | goto ret0; |
1398 | |
1399 | if (*s == '0') |
1400 | { |
1401 | nz0= 1; |
1402 | while (++s < end && *s == '0') ; |
1403 | if (s >= end) |
1404 | goto ret; |
1405 | } |
1406 | s0= s; |
1407 | y= z= 0; |
1408 | for (nd= nf= 0; s < end && (c= *s) >= '0' && c <= '9'; nd++, s++) |
1409 | if (nd < 9) |
1410 | y= 10*y + c - '0'; |
1411 | else if (nd < 16) |
1412 | z= 10*z + c - '0'; |
1413 | nd0= nd; |
1414 | if (s < end && c == '.') |
1415 | { |
1416 | ++s; |
1417 | if (!nd) |
1418 | { |
1419 | for (; s < end && (c= *s) == '0'; ++s) |
1420 | nz++; |
1421 | if (s < end && (c= *s) > '0' && c <= '9') |
1422 | { |
1423 | s0= s; |
1424 | nf+= nz; |
1425 | nz= 0; |
1426 | goto have_dig; |
1427 | } |
1428 | goto dig_done; |
1429 | } |
1430 | for (; s < end && (c= *s) >= '0' && c <= '9'; ++s) |
1431 | { |
1432 | have_dig: |
1433 | /* |
1434 | Here we are parsing the fractional part. |
1435 | We can stop counting digits after a while: the extra digits |
1436 | will not contribute to the actual result produced by s2b(). |
1437 | We have to continue scanning, in case there is an exponent part. |
1438 | */ |
1439 | if (nd < 2 * DBL_DIG) |
1440 | { |
1441 | nz++; |
1442 | if (c-= '0') |
1443 | { |
1444 | nf+= nz; |
1445 | for (i= 1; i < nz; i++) |
1446 | if (nd++ < 9) |
1447 | y*= 10; |
1448 | else if (nd <= DBL_DIG + 1) |
1449 | z*= 10; |
1450 | if (nd++ < 9) |
1451 | y= 10*y + c; |
1452 | else if (nd <= DBL_DIG + 1) |
1453 | z= 10*z + c; |
1454 | nz= 0; |
1455 | } |
1456 | } |
1457 | } |
1458 | } |
1459 | dig_done: |
1460 | e= 0; |
1461 | if (s < end && (c == 'e' || c == 'E')) |
1462 | { |
1463 | if (!nd && !nz && !nz0) |
1464 | goto ret0; |
1465 | s00= s; |
1466 | esign= 0; |
1467 | if (++s < end) |
1468 | switch (c= *s) { |
1469 | case '-': esign= 1; |
1470 | /* fall through */ |
1471 | case '+': c= *++s; |
1472 | } |
1473 | if (s < end && c >= '0' && c <= '9') |
1474 | { |
1475 | while (s < end && c == '0') |
1476 | c= *++s; |
1477 | if (s < end && c > '0' && c <= '9') { |
1478 | L= c - '0'; |
1479 | s1= s; |
1480 | while (++s < end && (c= *s) >= '0' && c <= '9') |
1481 | L= 10*L + c - '0'; |
1482 | if (s - s1 > 8 || L > 19999) |
1483 | /* Avoid confusion from exponents |
1484 | * so large that e might overflow. |
1485 | */ |
1486 | e= 19999; /* safe for 16 bit ints */ |
1487 | else |
1488 | e= (int)L; |
1489 | if (esign) |
1490 | e= -e; |
1491 | } |
1492 | else |
1493 | e= 0; |
1494 | } |
1495 | else |
1496 | s= s00; |
1497 | } |
1498 | if (!nd) |
1499 | { |
1500 | if (!nz && !nz0) |
1501 | { |
1502 | ret0: |
1503 | s= s00; |
1504 | sign= 0; |
1505 | } |
1506 | goto ret; |
1507 | } |
1508 | e1= e -= nf; |
1509 | |
1510 | /* |
1511 | Now we have nd0 digits, starting at s0, followed by a |
1512 | decimal point, followed by nd-nd0 digits. The number we're |
1513 | after is the integer represented by those digits times |
1514 | 10**e |
1515 | */ |
1516 | |
1517 | if (!nd0) |
1518 | nd0= nd; |
1519 | k= nd < DBL_DIG + 1 ? nd : DBL_DIG + 1; |
1520 | dval(&rv)= y; |
1521 | if (k > 9) |
1522 | { |
1523 | #ifdef SET_INEXACT |
1524 | if (k > DBL_DIG) |
1525 | oldinexact = get_inexact(); |
1526 | #endif |
1527 | dval(&rv)= tens[k - 9] * dval(&rv) + z; |
1528 | } |
1529 | bd0= 0; |
1530 | if (nd <= DBL_DIG |
1531 | #ifndef Honor_FLT_ROUNDS |
1532 | && Flt_Rounds == 1 |
1533 | #endif |
1534 | ) |
1535 | { |
1536 | if (!e) |
1537 | goto ret; |
1538 | if (e > 0) |
1539 | { |
1540 | if (e <= Ten_pmax) |
1541 | { |
1542 | #ifdef Honor_FLT_ROUNDS |
1543 | /* round correctly FLT_ROUNDS = 2 or 3 */ |
1544 | if (sign) |
1545 | { |
1546 | rv.d= -rv.d; |
1547 | sign= 0; |
1548 | } |
1549 | #endif |
1550 | /* rv = */ rounded_product(dval(&rv), tens[e]); |
1551 | goto ret; |
1552 | } |
1553 | i= DBL_DIG - nd; |
1554 | if (e <= Ten_pmax + i) |
1555 | { |
1556 | /* |
1557 | A fancier test would sometimes let us do |
1558 | this for larger i values. |
1559 | */ |
1560 | #ifdef Honor_FLT_ROUNDS |
1561 | /* round correctly FLT_ROUNDS = 2 or 3 */ |
1562 | if (sign) |
1563 | { |
1564 | rv.d= -rv.d; |
1565 | sign= 0; |
1566 | } |
1567 | #endif |
1568 | e-= i; |
1569 | dval(&rv)*= tens[i]; |
1570 | /* rv = */ rounded_product(dval(&rv), tens[e]); |
1571 | goto ret; |
1572 | } |
1573 | } |
1574 | #ifndef Inaccurate_Divide |
1575 | else if (e >= -Ten_pmax) |
1576 | { |
1577 | #ifdef Honor_FLT_ROUNDS |
1578 | /* round correctly FLT_ROUNDS = 2 or 3 */ |
1579 | if (sign) |
1580 | { |
1581 | rv.d= -rv.d; |
1582 | sign= 0; |
1583 | } |
1584 | #endif |
1585 | /* rv = */ rounded_quotient(dval(&rv), tens[-e]); |
1586 | goto ret; |
1587 | } |
1588 | #endif |
1589 | } |
1590 | e1+= nd - k; |
1591 | |
1592 | #ifdef SET_INEXACT |
1593 | inexact= 1; |
1594 | if (k <= DBL_DIG) |
1595 | oldinexact= get_inexact(); |
1596 | #endif |
1597 | scale= 0; |
1598 | #ifdef Honor_FLT_ROUNDS |
1599 | if ((rounding= Flt_Rounds) >= 2) |
1600 | { |
1601 | if (sign) |
1602 | rounding= rounding == 2 ? 0 : 2; |
1603 | else |
1604 | if (rounding != 2) |
1605 | rounding= 0; |
1606 | } |
1607 | #endif |
1608 | |
1609 | /* Get starting approximation = rv * 10**e1 */ |
1610 | |
1611 | if (e1 > 0) |
1612 | { |
1613 | if ((i= e1 & 15)) |
1614 | dval(&rv)*= tens[i]; |
1615 | if (e1&= ~15) |
1616 | { |
1617 | if (e1 > DBL_MAX_10_EXP) |
1618 | { |
1619 | ovfl: |
1620 | *error= EOVERFLOW; |
1621 | /* Can't trust HUGE_VAL */ |
1622 | #ifdef Honor_FLT_ROUNDS |
1623 | switch (rounding) |
1624 | { |
1625 | case 0: /* toward 0 */ |
1626 | case 3: /* toward -infinity */ |
1627 | word0(&rv)= Big0; |
1628 | word1(&rv)= Big1; |
1629 | break; |
1630 | default: |
1631 | word0(&rv)= Exp_mask; |
1632 | word1(&rv)= 0; |
1633 | } |
1634 | #else /*Honor_FLT_ROUNDS*/ |
1635 | word0(&rv)= Exp_mask; |
1636 | word1(&rv)= 0; |
1637 | #endif /*Honor_FLT_ROUNDS*/ |
1638 | #ifdef SET_INEXACT |
1639 | /* set overflow bit */ |
1640 | dval(&rv0)= 1e300; |
1641 | dval(&rv0)*= dval(&rv0); |
1642 | #endif |
1643 | if (bd0) |
1644 | goto retfree; |
1645 | goto ret; |
1646 | } |
1647 | e1>>= 4; |
1648 | for(j= 0; e1 > 1; j++, e1>>= 1) |
1649 | if (e1 & 1) |
1650 | dval(&rv)*= bigtens[j]; |
1651 | /* The last multiplication could overflow. */ |
1652 | word0(&rv)-= P*Exp_msk1; |
1653 | dval(&rv)*= bigtens[j]; |
1654 | if ((z= word0(&rv) & Exp_mask) > Exp_msk1 * (DBL_MAX_EXP + Bias - P)) |
1655 | goto ovfl; |
1656 | if (z > Exp_msk1 * (DBL_MAX_EXP + Bias - 1 - P)) |
1657 | { |
1658 | /* set to largest number (Can't trust DBL_MAX) */ |
1659 | word0(&rv)= Big0; |
1660 | word1(&rv)= Big1; |
1661 | } |
1662 | else |
1663 | word0(&rv)+= P*Exp_msk1; |
1664 | } |
1665 | } |
1666 | else if (e1 < 0) |
1667 | { |
1668 | e1= -e1; |
1669 | if ((i= e1 & 15)) |
1670 | dval(&rv)/= tens[i]; |
1671 | if ((e1>>= 4)) |
1672 | { |
1673 | if (e1 >= 1 << n_bigtens) |
1674 | goto undfl; |
1675 | if (e1 & Scale_Bit) |
1676 | scale= 2 * P; |
1677 | for(j= 0; e1 > 0; j++, e1>>= 1) |
1678 | if (e1 & 1) |
1679 | dval(&rv)*= tinytens[j]; |
1680 | if (scale && (j = 2 * P + 1 - ((word0(&rv) & Exp_mask) >> Exp_shift)) > 0) |
1681 | { |
1682 | /* scaled rv is denormal; zap j low bits */ |
1683 | if (j >= 32) |
1684 | { |
1685 | word1(&rv)= 0; |
1686 | if (j >= 53) |
1687 | word0(&rv)= (P + 2) * Exp_msk1; |
1688 | else |
1689 | word0(&rv)&= 0xffffffff << (j - 32); |
1690 | } |
1691 | else |
1692 | word1(&rv)&= 0xffffffff << j; |
1693 | } |
1694 | if (!dval(&rv)) |
1695 | { |
1696 | undfl: |
1697 | dval(&rv)= 0.; |
1698 | if (bd0) |
1699 | goto retfree; |
1700 | goto ret; |
1701 | } |
1702 | } |
1703 | } |
1704 | |
1705 | /* Now the hard part -- adjusting rv to the correct value.*/ |
1706 | |
1707 | /* Put digits into bd: true value = bd * 10^e */ |
1708 | |
1709 | bd0= s2b(s0, nd0, nd, y, &alloc); |
1710 | |
1711 | for(;;) |
1712 | { |
1713 | bd= Balloc(bd0->k, &alloc); |
1714 | Bcopy(bd, bd0); |
1715 | bb= d2b(&rv, &bbe, &bbbits, &alloc); /* rv = bb * 2^bbe */ |
1716 | bs= i2b(1, &alloc); |
1717 | |
1718 | if (e >= 0) |
1719 | { |
1720 | bb2= bb5= 0; |
1721 | bd2= bd5= e; |
1722 | } |
1723 | else |
1724 | { |
1725 | bb2= bb5= -e; |
1726 | bd2= bd5= 0; |
1727 | } |
1728 | if (bbe >= 0) |
1729 | bb2+= bbe; |
1730 | else |
1731 | bd2-= bbe; |
1732 | bs2= bb2; |
1733 | #ifdef Honor_FLT_ROUNDS |
1734 | if (rounding != 1) |
1735 | bs2++; |
1736 | #endif |
1737 | j= bbe - scale; |
1738 | i= j + bbbits - 1; /* logb(rv) */ |
1739 | if (i < Emin) /* denormal */ |
1740 | j+= P - Emin; |
1741 | else |
1742 | j= P + 1 - bbbits; |
1743 | bb2+= j; |
1744 | bd2+= j; |
1745 | bd2+= scale; |
1746 | i= bb2 < bd2 ? bb2 : bd2; |
1747 | if (i > bs2) |
1748 | i= bs2; |
1749 | if (i > 0) |
1750 | { |
1751 | bb2-= i; |
1752 | bd2-= i; |
1753 | bs2-= i; |
1754 | } |
1755 | if (bb5 > 0) |
1756 | { |
1757 | bs= pow5mult(bs, bb5, &alloc); |
1758 | bb1= mult(bs, bb, &alloc); |
1759 | Bfree(bb, &alloc); |
1760 | bb= bb1; |
1761 | } |
1762 | if (bb2 > 0) |
1763 | bb= lshift(bb, bb2, &alloc); |
1764 | if (bd5 > 0) |
1765 | bd= pow5mult(bd, bd5, &alloc); |
1766 | if (bd2 > 0) |
1767 | bd= lshift(bd, bd2, &alloc); |
1768 | if (bs2 > 0) |
1769 | bs= lshift(bs, bs2, &alloc); |
1770 | delta= diff(bb, bd, &alloc); |
1771 | dsign= delta->sign; |
1772 | delta->sign= 0; |
1773 | i= cmp(delta, bs); |
1774 | #ifdef Honor_FLT_ROUNDS |
1775 | if (rounding != 1) |
1776 | { |
1777 | if (i < 0) |
1778 | { |
1779 | /* Error is less than an ulp */ |
1780 | if (!delta->p.x[0] && delta->wds <= 1) |
1781 | { |
1782 | /* exact */ |
1783 | #ifdef SET_INEXACT |
1784 | inexact= 0; |
1785 | #endif |
1786 | break; |
1787 | } |
1788 | if (rounding) |
1789 | { |
1790 | if (dsign) |
1791 | { |
1792 | adj.d= 1.; |
1793 | goto apply_adj; |
1794 | } |
1795 | } |
1796 | else if (!dsign) |
1797 | { |
1798 | adj.d= -1.; |
1799 | if (!word1(&rv) && !(word0(&rv) & Frac_mask)) |
1800 | { |
1801 | y= word0(&rv) & Exp_mask; |
1802 | if (!scale || y > 2*P*Exp_msk1) |
1803 | { |
1804 | delta= lshift(delta, Log2P, &alloc); |
1805 | if (cmp(delta, bs) <= 0) |
1806 | adj.d= -0.5; |
1807 | } |
1808 | } |
1809 | apply_adj: |
1810 | if (scale && (y= word0(&rv) & Exp_mask) <= 2 * P * Exp_msk1) |
1811 | word0(&adj)+= (2 * P + 1) * Exp_msk1 - y; |
1812 | dval(&rv)+= adj.d * ulp(&rv); |
1813 | } |
1814 | break; |
1815 | } |
1816 | adj.d= ratio(delta, bs); |
1817 | if (adj.d < 1.) |
1818 | adj.d= 1.; |
1819 | if (adj.d <= 0x7ffffffe) |
1820 | { |
1821 | /* adj = rounding ? ceil(adj) : floor(adj); */ |
1822 | y= adj.d; |
1823 | if (y != adj.d) |
1824 | { |
1825 | if (!((rounding >> 1) ^ dsign)) |
1826 | y++; |
1827 | adj.d= y; |
1828 | } |
1829 | } |
1830 | if (scale && (y= word0(&rv) & Exp_mask) <= 2 * P * Exp_msk1) |
1831 | word0(&adj)+= (2 * P + 1) * Exp_msk1 - y; |
1832 | adj.d*= ulp(&rv); |
1833 | if (dsign) |
1834 | dval(&rv)+= adj.d; |
1835 | else |
1836 | dval(&rv)-= adj.d; |
1837 | goto cont; |
1838 | } |
1839 | #endif /*Honor_FLT_ROUNDS*/ |
1840 | |
1841 | if (i < 0) |
1842 | { |
1843 | /* |
1844 | Error is less than half an ulp -- check for special case of mantissa |
1845 | a power of two. |
1846 | */ |
1847 | if (dsign || word1(&rv) || word0(&rv) & Bndry_mask || |
1848 | (word0(&rv) & Exp_mask) <= (2 * P + 1) * Exp_msk1) |
1849 | { |
1850 | #ifdef SET_INEXACT |
1851 | if (!delta->x[0] && delta->wds <= 1) |
1852 | inexact= 0; |
1853 | #endif |
1854 | break; |
1855 | } |
1856 | if (!delta->p.x[0] && delta->wds <= 1) |
1857 | { |
1858 | /* exact result */ |
1859 | #ifdef SET_INEXACT |
1860 | inexact= 0; |
1861 | #endif |
1862 | break; |
1863 | } |
1864 | delta= lshift(delta, Log2P, &alloc); |
1865 | if (cmp(delta, bs) > 0) |
1866 | goto drop_down; |
1867 | break; |
1868 | } |
1869 | if (i == 0) |
1870 | { |
1871 | /* exactly half-way between */ |
1872 | if (dsign) |
1873 | { |
1874 | if ((word0(&rv) & Bndry_mask1) == Bndry_mask1 && |
1875 | word1(&rv) == |
1876 | ((scale && (y = word0(&rv) & Exp_mask) <= 2 * P * Exp_msk1) ? |
1877 | (0xffffffff & (0xffffffff << (2*P+1-(y>>Exp_shift)))) : |
1878 | 0xffffffff)) |
1879 | { |
1880 | /*boundary case -- increment exponent*/ |
1881 | word0(&rv)= (word0(&rv) & Exp_mask) + Exp_msk1; |
1882 | word1(&rv) = 0; |
1883 | dsign = 0; |
1884 | break; |
1885 | } |
1886 | } |
1887 | else if (!(word0(&rv) & Bndry_mask) && !word1(&rv)) |
1888 | { |
1889 | drop_down: |
1890 | /* boundary case -- decrement exponent */ |
1891 | if (scale) |
1892 | { |
1893 | L= word0(&rv) & Exp_mask; |
1894 | if (L <= (2 *P + 1) * Exp_msk1) |
1895 | { |
1896 | if (L > (P + 2) * Exp_msk1) |
1897 | /* round even ==> accept rv */ |
1898 | break; |
1899 | /* rv = smallest denormal */ |
1900 | goto undfl; |
1901 | } |
1902 | } |
1903 | L= (word0(&rv) & Exp_mask) - Exp_msk1; |
1904 | word0(&rv)= L | Bndry_mask1; |
1905 | word1(&rv)= 0xffffffff; |
1906 | break; |
1907 | } |
1908 | if (!(word1(&rv) & LSB)) |
1909 | break; |
1910 | if (dsign) |
1911 | dval(&rv)+= ulp(&rv); |
1912 | else |
1913 | { |
1914 | dval(&rv)-= ulp(&rv); |
1915 | if (!dval(&rv)) |
1916 | goto undfl; |
1917 | } |
1918 | dsign= 1 - dsign; |
1919 | break; |
1920 | } |
1921 | if ((aadj= ratio(delta, bs)) <= 2.) |
1922 | { |
1923 | if (dsign) |
1924 | aadj= aadj1= 1.; |
1925 | else if (word1(&rv) || word0(&rv) & Bndry_mask) |
1926 | { |
1927 | if (word1(&rv) == Tiny1 && !word0(&rv)) |
1928 | goto undfl; |
1929 | aadj= 1.; |
1930 | aadj1= -1.; |
1931 | } |
1932 | else |
1933 | { |
1934 | /* special case -- power of FLT_RADIX to be rounded down... */ |
1935 | if (aadj < 2. / FLT_RADIX) |
1936 | aadj= 1. / FLT_RADIX; |
1937 | else |
1938 | aadj*= 0.5; |
1939 | aadj1= -aadj; |
1940 | } |
1941 | } |
1942 | else |
1943 | { |
1944 | aadj*= 0.5; |
1945 | aadj1= dsign ? aadj : -aadj; |
1946 | #ifdef Check_FLT_ROUNDS |
1947 | switch (Rounding) |
1948 | { |
1949 | case 2: /* towards +infinity */ |
1950 | aadj1-= 0.5; |
1951 | break; |
1952 | case 0: /* towards 0 */ |
1953 | case 3: /* towards -infinity */ |
1954 | aadj1+= 0.5; |
1955 | } |
1956 | #else |
1957 | if (Flt_Rounds == 0) |
1958 | aadj1+= 0.5; |
1959 | #endif /*Check_FLT_ROUNDS*/ |
1960 | } |
1961 | y= word0(&rv) & Exp_mask; |
1962 | |
1963 | /* Check for overflow */ |
1964 | |
1965 | if (y == Exp_msk1 * (DBL_MAX_EXP + Bias - 1)) |
1966 | { |
1967 | dval(&rv0)= dval(&rv); |
1968 | word0(&rv)-= P * Exp_msk1; |
1969 | adj.d= aadj1 * ulp(&rv); |
1970 | dval(&rv)+= adj.d; |
1971 | if ((word0(&rv) & Exp_mask) >= Exp_msk1 * (DBL_MAX_EXP + Bias - P)) |
1972 | { |
1973 | if (word0(&rv0) == Big0 && word1(&rv0) == Big1) |
1974 | goto ovfl; |
1975 | word0(&rv)= Big0; |
1976 | word1(&rv)= Big1; |
1977 | goto cont; |
1978 | } |
1979 | else |
1980 | word0(&rv)+= P * Exp_msk1; |
1981 | } |
1982 | else |
1983 | { |
1984 | if (scale && y <= 2 * P * Exp_msk1) |
1985 | { |
1986 | if (aadj <= 0x7fffffff) |
1987 | { |
1988 | if ((z= (ULong) aadj) <= 0) |
1989 | z= 1; |
1990 | aadj= z; |
1991 | aadj1= dsign ? aadj : -aadj; |
1992 | } |
1993 | dval(&aadj2) = aadj1; |
1994 | word0(&aadj2)+= (2 * P + 1) * Exp_msk1 - y; |
1995 | aadj1= dval(&aadj2); |
1996 | adj.d= aadj1 * ulp(&rv); |
1997 | dval(&rv)+= adj.d; |
1998 | if (rv.d == 0.) |
1999 | goto undfl; |
2000 | } |
2001 | else |
2002 | { |
2003 | adj.d= aadj1 * ulp(&rv); |
2004 | dval(&rv)+= adj.d; |
2005 | } |
2006 | } |
2007 | z= word0(&rv) & Exp_mask; |
2008 | #ifndef SET_INEXACT |
2009 | if (!scale) |
2010 | if (y == z) |
2011 | { |
2012 | /* Can we stop now? */ |
2013 | L= (Long)aadj; |
2014 | aadj-= L; |
2015 | /* The tolerances below are conservative. */ |
2016 | if (dsign || word1(&rv) || word0(&rv) & Bndry_mask) |
2017 | { |
2018 | if (aadj < .4999999 || aadj > .5000001) |
2019 | break; |
2020 | } |
2021 | else if (aadj < .4999999 / FLT_RADIX) |
2022 | break; |
2023 | } |
2024 | #endif |
2025 | cont: |
2026 | Bfree(bb, &alloc); |
2027 | Bfree(bd, &alloc); |
2028 | Bfree(bs, &alloc); |
2029 | Bfree(delta, &alloc); |
2030 | } |
2031 | #ifdef SET_INEXACT |
2032 | if (inexact) |
2033 | { |
2034 | if (!oldinexact) |
2035 | { |
2036 | word0(&rv0)= Exp_1 + (70 << Exp_shift); |
2037 | word1(&rv0)= 0; |
2038 | dval(&rv0)+= 1.; |
2039 | } |
2040 | } |
2041 | else if (!oldinexact) |
2042 | clear_inexact(); |
2043 | #endif |
2044 | if (scale) |
2045 | { |
2046 | word0(&rv0)= Exp_1 - 2 * P * Exp_msk1; |
2047 | word1(&rv0)= 0; |
2048 | dval(&rv)*= dval(&rv0); |
2049 | } |
2050 | #ifdef SET_INEXACT |
2051 | if (inexact && !(word0(&rv) & Exp_mask)) |
2052 | { |
2053 | /* set underflow bit */ |
2054 | dval(&rv0)= 1e-300; |
2055 | dval(&rv0)*= dval(&rv0); |
2056 | } |
2057 | #endif |
2058 | retfree: |
2059 | Bfree(bb, &alloc); |
2060 | Bfree(bd, &alloc); |
2061 | Bfree(bs, &alloc); |
2062 | Bfree(bd0, &alloc); |
2063 | Bfree(delta, &alloc); |
2064 | ret: |
2065 | *se= (char *)s; |
2066 | return sign ? -dval(&rv) : dval(&rv); |
2067 | } |
2068 | |
2069 | |
2070 | static int quorem(Bigint *b, Bigint *S) |
2071 | { |
2072 | int n; |
2073 | ULong *bx, *bxe, q, *sx, *sxe; |
2074 | ULLong borrow, carry, y, ys; |
2075 | |
2076 | n= S->wds; |
2077 | if (b->wds < n) |
2078 | return 0; |
2079 | sx= S->p.x; |
2080 | sxe= sx + --n; |
2081 | bx= b->p.x; |
2082 | bxe= bx + n; |
2083 | q= *bxe / (*sxe + 1); /* ensure q <= true quotient */ |
2084 | if (q) |
2085 | { |
2086 | borrow= 0; |
2087 | carry= 0; |
2088 | do |
2089 | { |
2090 | ys= *sx++ * (ULLong)q + carry; |
2091 | carry= ys >> 32; |
2092 | y= *bx - (ys & FFFFFFFF) - borrow; |
2093 | borrow= y >> 32 & (ULong)1; |
2094 | *bx++= (ULong) (y & FFFFFFFF); |
2095 | } |
2096 | while (sx <= sxe); |
2097 | if (!*bxe) |
2098 | { |
2099 | bx= b->p.x; |
2100 | while (--bxe > bx && !*bxe) |
2101 | --n; |
2102 | b->wds= n; |
2103 | } |
2104 | } |
2105 | if (cmp(b, S) >= 0) |
2106 | { |
2107 | q++; |
2108 | borrow= 0; |
2109 | carry= 0; |
2110 | bx= b->p.x; |
2111 | sx= S->p.x; |
2112 | do |
2113 | { |
2114 | ys= *sx++ + carry; |
2115 | carry= ys >> 32; |
2116 | y= *bx - (ys & FFFFFFFF) - borrow; |
2117 | borrow= y >> 32 & (ULong)1; |
2118 | *bx++= (ULong) (y & FFFFFFFF); |
2119 | } |
2120 | while (sx <= sxe); |
2121 | bx= b->p.x; |
2122 | bxe= bx + n; |
2123 | if (!*bxe) |
2124 | { |
2125 | while (--bxe > bx && !*bxe) |
2126 | --n; |
2127 | b->wds= n; |
2128 | } |
2129 | } |
2130 | return q; |
2131 | } |
2132 | |
2133 | |
2134 | /* |
2135 | dtoa for IEEE arithmetic (dmg): convert double to ASCII string. |
2136 | |
2137 | Inspired by "How to Print Floating-Point Numbers Accurately" by |
2138 | Guy L. Steele, Jr. and Jon L. White [Proc. ACM SIGPLAN '90, pp. 112-126]. |
2139 | |
2140 | Modifications: |
2141 | 1. Rather than iterating, we use a simple numeric overestimate |
2142 | to determine k= floor(log10(d)). We scale relevant |
2143 | quantities using O(log2(k)) rather than O(k) multiplications. |
2144 | 2. For some modes > 2 (corresponding to ecvt and fcvt), we don't |
2145 | try to generate digits strictly left to right. Instead, we |
2146 | compute with fewer bits and propagate the carry if necessary |
2147 | when rounding the final digit up. This is often faster. |
2148 | 3. Under the assumption that input will be rounded nearest, |
2149 | mode 0 renders 1e23 as 1e23 rather than 9.999999999999999e22. |
2150 | That is, we allow equality in stopping tests when the |
2151 | round-nearest rule will give the same floating-point value |
2152 | as would satisfaction of the stopping test with strict |
2153 | inequality. |
2154 | 4. We remove common factors of powers of 2 from relevant |
2155 | quantities. |
2156 | 5. When converting floating-point integers less than 1e16, |
2157 | we use floating-point arithmetic rather than resorting |
2158 | to multiple-precision integers. |
2159 | 6. When asked to produce fewer than 15 digits, we first try |
2160 | to get by with floating-point arithmetic; we resort to |
2161 | multiple-precision integer arithmetic only if we cannot |
2162 | guarantee that the floating-point calculation has given |
2163 | the correctly rounded result. For k requested digits and |
2164 | "uniformly" distributed input, the probability is |
2165 | something like 10^(k-15) that we must resort to the Long |
2166 | calculation. |
2167 | */ |
2168 | |
2169 | static char *dtoa(double dd, int mode, int ndigits, int *decpt, int *sign, |
2170 | char **rve, char *buf, size_t buf_size) |
2171 | { |
2172 | /* |
2173 | Arguments ndigits, decpt, sign are similar to those |
2174 | of ecvt and fcvt; trailing zeros are suppressed from |
2175 | the returned string. If not null, *rve is set to point |
2176 | to the end of the return value. If d is +-Infinity or NaN, |
2177 | then *decpt is set to DTOA_OVERFLOW. |
2178 | |
2179 | mode: |
2180 | 0 ==> shortest string that yields d when read in |
2181 | and rounded to nearest. |
2182 | 1 ==> like 0, but with Steele & White stopping rule; |
2183 | e.g. with IEEE P754 arithmetic , mode 0 gives |
2184 | 1e23 whereas mode 1 gives 9.999999999999999e22. |
2185 | 2 ==> MY_MAX(1,ndigits) significant digits. This gives a |
2186 | return value similar to that of ecvt, except |
2187 | that trailing zeros are suppressed. |
2188 | 3 ==> through ndigits past the decimal point. This |
2189 | gives a return value similar to that from fcvt, |
2190 | except that trailing zeros are suppressed, and |
2191 | ndigits can be negative. |
2192 | 4,5 ==> similar to 2 and 3, respectively, but (in |
2193 | round-nearest mode) with the tests of mode 0 to |
2194 | possibly return a shorter string that rounds to d. |
2195 | With IEEE arithmetic and compilation with |
2196 | -DHonor_FLT_ROUNDS, modes 4 and 5 behave the same |
2197 | as modes 2 and 3 when FLT_ROUNDS != 1. |
2198 | 6-9 ==> Debugging modes similar to mode - 4: don't try |
2199 | fast floating-point estimate (if applicable). |
2200 | |
2201 | Values of mode other than 0-9 are treated as mode 0. |
2202 | |
2203 | Sufficient space is allocated to the return value |
2204 | to hold the suppressed trailing zeros. |
2205 | */ |
2206 | |
2207 | int bbits, b2, b5, be, dig, i, ieps, UNINIT_VAR(ilim), ilim0, |
2208 | UNINIT_VAR(ilim1), j, j1, k, k0, k_check, leftright, m2, m5, s2, s5, |
2209 | spec_case, try_quick; |
2210 | Long L; |
2211 | int denorm; |
2212 | ULong x; |
2213 | Bigint *b, *b1, *delta, *mlo, *mhi, *S; |
2214 | U d2, eps, u; |
2215 | double ds; |
2216 | char *s, *s0; |
2217 | #ifdef Honor_FLT_ROUNDS |
2218 | int rounding; |
2219 | #endif |
2220 | Stack_alloc alloc; |
2221 | |
2222 | alloc.begin= alloc.free= buf; |
2223 | alloc.end= buf + buf_size; |
2224 | memset(alloc.freelist, 0, sizeof(alloc.freelist)); |
2225 | |
2226 | u.d= dd; |
2227 | if (word0(&u) & Sign_bit) |
2228 | { |
2229 | /* set sign for everything, including 0's and NaNs */ |
2230 | *sign= 1; |
2231 | word0(&u) &= ~Sign_bit; /* clear sign bit */ |
2232 | } |
2233 | else |
2234 | *sign= 0; |
2235 | |
2236 | /* If infinity, set decpt to DTOA_OVERFLOW, if 0 set it to 1 */ |
2237 | if (((word0(&u) & Exp_mask) == Exp_mask && (*decpt= DTOA_OVERFLOW)) || |
2238 | (!dval(&u) && (*decpt= 1))) |
2239 | { |
2240 | /* Infinity, NaN, 0 */ |
2241 | char *res= (char*) dtoa_alloc(2, &alloc); |
2242 | res[0]= '0'; |
2243 | res[1]= '\0'; |
2244 | if (rve) |
2245 | *rve= res + 1; |
2246 | return res; |
2247 | } |
2248 | |
2249 | #ifdef Honor_FLT_ROUNDS |
2250 | if ((rounding= Flt_Rounds) >= 2) |
2251 | { |
2252 | if (*sign) |
2253 | rounding= rounding == 2 ? 0 : 2; |
2254 | else |
2255 | if (rounding != 2) |
2256 | rounding= 0; |
2257 | } |
2258 | #endif |
2259 | |
2260 | b= d2b(&u, &be, &bbits, &alloc); |
2261 | if ((i= (int)(word0(&u) >> Exp_shift1 & (Exp_mask>>Exp_shift1)))) |
2262 | { |
2263 | dval(&d2)= dval(&u); |
2264 | word0(&d2) &= Frac_mask1; |
2265 | word0(&d2) |= Exp_11; |
2266 | |
2267 | /* |
2268 | log(x) ~=~ log(1.5) + (x-1.5)/1.5 |
2269 | log10(x) = log(x) / log(10) |
2270 | ~=~ log(1.5)/log(10) + (x-1.5)/(1.5*log(10)) |
2271 | log10(d)= (i-Bias)*log(2)/log(10) + log10(d2) |
2272 | |
2273 | This suggests computing an approximation k to log10(d) by |
2274 | |
2275 | k= (i - Bias)*0.301029995663981 |
2276 | + ( (d2-1.5)*0.289529654602168 + 0.176091259055681 ); |
2277 | |
2278 | We want k to be too large rather than too small. |
2279 | The error in the first-order Taylor series approximation |
2280 | is in our favor, so we just round up the constant enough |
2281 | to compensate for any error in the multiplication of |
2282 | (i - Bias) by 0.301029995663981; since |i - Bias| <= 1077, |
2283 | and 1077 * 0.30103 * 2^-52 ~=~ 7.2e-14, |
2284 | adding 1e-13 to the constant term more than suffices. |
2285 | Hence we adjust the constant term to 0.1760912590558. |
2286 | (We could get a more accurate k by invoking log10, |
2287 | but this is probably not worthwhile.) |
2288 | */ |
2289 | |
2290 | i-= Bias; |
2291 | denorm= 0; |
2292 | } |
2293 | else |
2294 | { |
2295 | /* d is denormalized */ |
2296 | |
2297 | i= bbits + be + (Bias + (P-1) - 1); |
2298 | x= i > 32 ? word0(&u) << (64 - i) | word1(&u) >> (i - 32) |
2299 | : word1(&u) << (32 - i); |
2300 | dval(&d2)= x; |
2301 | word0(&d2)-= 31*Exp_msk1; /* adjust exponent */ |
2302 | i-= (Bias + (P-1) - 1) + 1; |
2303 | denorm= 1; |
2304 | } |
2305 | ds= (dval(&d2)-1.5)*0.289529654602168 + 0.1760912590558 + i*0.301029995663981; |
2306 | k= (int)ds; |
2307 | if (ds < 0. && ds != k) |
2308 | k--; /* want k= floor(ds) */ |
2309 | k_check= 1; |
2310 | if (k >= 0 && k <= Ten_pmax) |
2311 | { |
2312 | if (dval(&u) < tens[k]) |
2313 | k--; |
2314 | k_check= 0; |
2315 | } |
2316 | j= bbits - i - 1; |
2317 | if (j >= 0) |
2318 | { |
2319 | b2= 0; |
2320 | s2= j; |
2321 | } |
2322 | else |
2323 | { |
2324 | b2= -j; |
2325 | s2= 0; |
2326 | } |
2327 | if (k >= 0) |
2328 | { |
2329 | b5= 0; |
2330 | s5= k; |
2331 | s2+= k; |
2332 | } |
2333 | else |
2334 | { |
2335 | b2-= k; |
2336 | b5= -k; |
2337 | s5= 0; |
2338 | } |
2339 | if (mode < 0 || mode > 9) |
2340 | mode= 0; |
2341 | |
2342 | #ifdef Check_FLT_ROUNDS |
2343 | try_quick= Rounding == 1; |
2344 | #else |
2345 | try_quick= 1; |
2346 | #endif |
2347 | |
2348 | if (mode > 5) |
2349 | { |
2350 | mode-= 4; |
2351 | try_quick= 0; |
2352 | } |
2353 | leftright= 1; |
2354 | switch (mode) { |
2355 | case 0: |
2356 | case 1: |
2357 | ilim= ilim1= -1; |
2358 | i= 18; |
2359 | ndigits= 0; |
2360 | break; |
2361 | case 2: |
2362 | leftright= 0; |
2363 | /* fall through */ |
2364 | case 4: |
2365 | if (ndigits <= 0) |
2366 | ndigits= 1; |
2367 | ilim= ilim1= i= ndigits; |
2368 | break; |
2369 | case 3: |
2370 | leftright= 0; |
2371 | /* fall through */ |
2372 | case 5: |
2373 | i= ndigits + k + 1; |
2374 | ilim= i; |
2375 | ilim1= i - 1; |
2376 | if (i <= 0) |
2377 | i= 1; |
2378 | } |
2379 | s= s0= dtoa_alloc(i, &alloc); |
2380 | |
2381 | #ifdef Honor_FLT_ROUNDS |
2382 | if (mode > 1 && rounding != 1) |
2383 | leftright= 0; |
2384 | #endif |
2385 | |
2386 | if (ilim >= 0 && ilim <= Quick_max && try_quick) |
2387 | { |
2388 | /* Try to get by with floating-point arithmetic. */ |
2389 | i= 0; |
2390 | dval(&d2)= dval(&u); |
2391 | k0= k; |
2392 | ilim0= ilim; |
2393 | ieps= 2; /* conservative */ |
2394 | if (k > 0) |
2395 | { |
2396 | ds= tens[k&0xf]; |
2397 | j= k >> 4; |
2398 | if (j & Bletch) |
2399 | { |
2400 | /* prevent overflows */ |
2401 | j&= Bletch - 1; |
2402 | dval(&u)/= bigtens[n_bigtens-1]; |
2403 | ieps++; |
2404 | } |
2405 | for (; j; j>>= 1, i++) |
2406 | { |
2407 | if (j & 1) |
2408 | { |
2409 | ieps++; |
2410 | ds*= bigtens[i]; |
2411 | } |
2412 | } |
2413 | dval(&u)/= ds; |
2414 | } |
2415 | else if ((j1= -k)) |
2416 | { |
2417 | dval(&u)*= tens[j1 & 0xf]; |
2418 | for (j= j1 >> 4; j; j>>= 1, i++) |
2419 | { |
2420 | if (j & 1) |
2421 | { |
2422 | ieps++; |
2423 | dval(&u)*= bigtens[i]; |
2424 | } |
2425 | } |
2426 | } |
2427 | if (k_check && dval(&u) < 1. && ilim > 0) |
2428 | { |
2429 | if (ilim1 <= 0) |
2430 | goto fast_failed; |
2431 | ilim= ilim1; |
2432 | k--; |
2433 | dval(&u)*= 10.; |
2434 | ieps++; |
2435 | } |
2436 | dval(&eps)= ieps*dval(&u) + 7.; |
2437 | word0(&eps)-= (P-1)*Exp_msk1; |
2438 | if (ilim == 0) |
2439 | { |
2440 | S= mhi= 0; |
2441 | dval(&u)-= 5.; |
2442 | if (dval(&u) > dval(&eps)) |
2443 | goto one_digit; |
2444 | if (dval(&u) < -dval(&eps)) |
2445 | goto no_digits; |
2446 | goto fast_failed; |
2447 | } |
2448 | if (leftright) |
2449 | { |
2450 | /* Use Steele & White method of only generating digits needed. */ |
2451 | dval(&eps)= 0.5/tens[ilim-1] - dval(&eps); |
2452 | for (i= 0;;) |
2453 | { |
2454 | L= (Long) dval(&u); |
2455 | dval(&u)-= L; |
2456 | *s++= '0' + (int)L; |
2457 | if (dval(&u) < dval(&eps)) |
2458 | goto ret1; |
2459 | if (1. - dval(&u) < dval(&eps)) |
2460 | goto bump_up; |
2461 | if (++i >= ilim) |
2462 | break; |
2463 | dval(&eps)*= 10.; |
2464 | dval(&u)*= 10.; |
2465 | } |
2466 | } |
2467 | else |
2468 | { |
2469 | /* Generate ilim digits, then fix them up. */ |
2470 | dval(&eps)*= tens[ilim-1]; |
2471 | for (i= 1;; i++, dval(&u)*= 10.) |
2472 | { |
2473 | L= (Long)(dval(&u)); |
2474 | if (!(dval(&u)-= L)) |
2475 | ilim= i; |
2476 | *s++= '0' + (int)L; |
2477 | if (i == ilim) |
2478 | { |
2479 | if (dval(&u) > 0.5 + dval(&eps)) |
2480 | goto bump_up; |
2481 | else if (dval(&u) < 0.5 - dval(&eps)) |
2482 | { |
2483 | while (*--s == '0'); |
2484 | s++; |
2485 | goto ret1; |
2486 | } |
2487 | break; |
2488 | } |
2489 | } |
2490 | } |
2491 | fast_failed: |
2492 | s= s0; |
2493 | dval(&u)= dval(&d2); |
2494 | k= k0; |
2495 | ilim= ilim0; |
2496 | } |
2497 | |
2498 | /* Do we have a "small" integer? */ |
2499 | |
2500 | if (be >= 0 && k <= Int_max) |
2501 | { |
2502 | /* Yes. */ |
2503 | ds= tens[k]; |
2504 | if (ndigits < 0 && ilim <= 0) |
2505 | { |
2506 | S= mhi= 0; |
2507 | if (ilim < 0 || dval(&u) <= 5*ds) |
2508 | goto no_digits; |
2509 | goto one_digit; |
2510 | } |
2511 | for (i= 1;; i++, dval(&u)*= 10.) |
2512 | { |
2513 | L= (Long)(dval(&u) / ds); |
2514 | dval(&u)-= L*ds; |
2515 | #ifdef Check_FLT_ROUNDS |
2516 | /* If FLT_ROUNDS == 2, L will usually be high by 1 */ |
2517 | if (dval(&u) < 0) |
2518 | { |
2519 | L--; |
2520 | dval(&u)+= ds; |
2521 | } |
2522 | #endif |
2523 | *s++= '0' + (int)L; |
2524 | if (!dval(&u)) |
2525 | { |
2526 | break; |
2527 | } |
2528 | if (i == ilim) |
2529 | { |
2530 | #ifdef Honor_FLT_ROUNDS |
2531 | if (mode > 1) |
2532 | { |
2533 | switch (rounding) { |
2534 | case 0: goto ret1; |
2535 | case 2: goto bump_up; |
2536 | } |
2537 | } |
2538 | #endif |
2539 | dval(&u)+= dval(&u); |
2540 | if (dval(&u) > ds || (dval(&u) == ds && L & 1)) |
2541 | { |
2542 | bump_up: |
2543 | while (*--s == '9') |
2544 | if (s == s0) |
2545 | { |
2546 | k++; |
2547 | *s= '0'; |
2548 | break; |
2549 | } |
2550 | ++*s++; |
2551 | } |
2552 | break; |
2553 | } |
2554 | } |
2555 | goto ret1; |
2556 | } |
2557 | |
2558 | m2= b2; |
2559 | m5= b5; |
2560 | mhi= mlo= 0; |
2561 | if (leftright) |
2562 | { |
2563 | i = denorm ? be + (Bias + (P-1) - 1 + 1) : 1 + P - bbits; |
2564 | b2+= i; |
2565 | s2+= i; |
2566 | mhi= i2b(1, &alloc); |
2567 | } |
2568 | if (m2 > 0 && s2 > 0) |
2569 | { |
2570 | i= m2 < s2 ? m2 : s2; |
2571 | b2-= i; |
2572 | m2-= i; |
2573 | s2-= i; |
2574 | } |
2575 | if (b5 > 0) |
2576 | { |
2577 | if (leftright) |
2578 | { |
2579 | if (m5 > 0) |
2580 | { |
2581 | mhi= pow5mult(mhi, m5, &alloc); |
2582 | b1= mult(mhi, b, &alloc); |
2583 | Bfree(b, &alloc); |
2584 | b= b1; |
2585 | } |
2586 | if ((j= b5 - m5)) |
2587 | b= pow5mult(b, j, &alloc); |
2588 | } |
2589 | else |
2590 | b= pow5mult(b, b5, &alloc); |
2591 | } |
2592 | S= i2b(1, &alloc); |
2593 | if (s5 > 0) |
2594 | S= pow5mult(S, s5, &alloc); |
2595 | |
2596 | /* Check for special case that d is a normalized power of 2. */ |
2597 | |
2598 | spec_case= 0; |
2599 | if ((mode < 2 || leftright) |
2600 | #ifdef Honor_FLT_ROUNDS |
2601 | && rounding == 1 |
2602 | #endif |
2603 | ) |
2604 | { |
2605 | if (!word1(&u) && !(word0(&u) & Bndry_mask) && |
2606 | word0(&u) & (Exp_mask & ~Exp_msk1) |
2607 | ) |
2608 | { |
2609 | /* The special case */ |
2610 | b2+= Log2P; |
2611 | s2+= Log2P; |
2612 | spec_case= 1; |
2613 | } |
2614 | } |
2615 | |
2616 | /* |
2617 | Arrange for convenient computation of quotients: |
2618 | shift left if necessary so divisor has 4 leading 0 bits. |
2619 | |
2620 | Perhaps we should just compute leading 28 bits of S once |
2621 | a nd for all and pass them and a shift to quorem, so it |
2622 | can do shifts and ors to compute the numerator for q. |
2623 | */ |
2624 | if ((i= ((s5 ? 32 - hi0bits(S->p.x[S->wds-1]) : 1) + s2) & 0x1f)) |
2625 | i= 32 - i; |
2626 | if (i > 4) |
2627 | { |
2628 | i-= 4; |
2629 | b2+= i; |
2630 | m2+= i; |
2631 | s2+= i; |
2632 | } |
2633 | else if (i < 4) |
2634 | { |
2635 | i+= 28; |
2636 | b2+= i; |
2637 | m2+= i; |
2638 | s2+= i; |
2639 | } |
2640 | if (b2 > 0) |
2641 | b= lshift(b, b2, &alloc); |
2642 | if (s2 > 0) |
2643 | S= lshift(S, s2, &alloc); |
2644 | if (k_check) |
2645 | { |
2646 | if (cmp(b,S) < 0) |
2647 | { |
2648 | k--; |
2649 | /* we botched the k estimate */ |
2650 | b= multadd(b, 10, 0, &alloc); |
2651 | if (leftright) |
2652 | mhi= multadd(mhi, 10, 0, &alloc); |
2653 | ilim= ilim1; |
2654 | } |
2655 | } |
2656 | if (ilim <= 0 && (mode == 3 || mode == 5)) |
2657 | { |
2658 | if (ilim < 0 || cmp(b,S= multadd(S,5,0, &alloc)) <= 0) |
2659 | { |
2660 | /* no digits, fcvt style */ |
2661 | no_digits: |
2662 | k= -1 - ndigits; |
2663 | goto ret; |
2664 | } |
2665 | one_digit: |
2666 | *s++= '1'; |
2667 | k++; |
2668 | goto ret; |
2669 | } |
2670 | if (leftright) |
2671 | { |
2672 | if (m2 > 0) |
2673 | mhi= lshift(mhi, m2, &alloc); |
2674 | |
2675 | /* |
2676 | Compute mlo -- check for special case that d is a normalized power of 2. |
2677 | */ |
2678 | |
2679 | mlo= mhi; |
2680 | if (spec_case) |
2681 | { |
2682 | mhi= Balloc(mhi->k, &alloc); |
2683 | Bcopy(mhi, mlo); |
2684 | mhi= lshift(mhi, Log2P, &alloc); |
2685 | } |
2686 | |
2687 | for (i= 1;;i++) |
2688 | { |
2689 | dig= quorem(b,S) + '0'; |
2690 | /* Do we yet have the shortest decimal string that will round to d? */ |
2691 | j= cmp(b, mlo); |
2692 | delta= diff(S, mhi, &alloc); |
2693 | j1= delta->sign ? 1 : cmp(b, delta); |
2694 | Bfree(delta, &alloc); |
2695 | if (j1 == 0 && mode != 1 && !(word1(&u) & 1) |
2696 | #ifdef Honor_FLT_ROUNDS |
2697 | && rounding >= 1 |
2698 | #endif |
2699 | ) |
2700 | { |
2701 | if (dig == '9') |
2702 | goto round_9_up; |
2703 | if (j > 0) |
2704 | dig++; |
2705 | *s++= dig; |
2706 | goto ret; |
2707 | } |
2708 | if (j < 0 || (j == 0 && mode != 1 && !(word1(&u) & 1))) |
2709 | { |
2710 | if (!b->p.x[0] && b->wds <= 1) |
2711 | { |
2712 | goto accept_dig; |
2713 | } |
2714 | #ifdef Honor_FLT_ROUNDS |
2715 | if (mode > 1) |
2716 | switch (rounding) { |
2717 | case 0: goto accept_dig; |
2718 | case 2: goto keep_dig; |
2719 | } |
2720 | #endif /*Honor_FLT_ROUNDS*/ |
2721 | if (j1 > 0) |
2722 | { |
2723 | b= lshift(b, 1, &alloc); |
2724 | j1= cmp(b, S); |
2725 | if ((j1 > 0 || (j1 == 0 && dig & 1)) |
2726 | && dig++ == '9') |
2727 | goto round_9_up; |
2728 | } |
2729 | accept_dig: |
2730 | *s++= dig; |
2731 | goto ret; |
2732 | } |
2733 | if (j1 > 0) |
2734 | { |
2735 | #ifdef Honor_FLT_ROUNDS |
2736 | if (!rounding) |
2737 | goto accept_dig; |
2738 | #endif |
2739 | if (dig == '9') |
2740 | { /* possible if i == 1 */ |
2741 | round_9_up: |
2742 | *s++= '9'; |
2743 | goto roundoff; |
2744 | } |
2745 | *s++= dig + 1; |
2746 | goto ret; |
2747 | } |
2748 | #ifdef Honor_FLT_ROUNDS |
2749 | keep_dig: |
2750 | #endif |
2751 | *s++= dig; |
2752 | if (i == ilim) |
2753 | break; |
2754 | b= multadd(b, 10, 0, &alloc); |
2755 | if (mlo == mhi) |
2756 | mlo= mhi= multadd(mhi, 10, 0, &alloc); |
2757 | else |
2758 | { |
2759 | mlo= multadd(mlo, 10, 0, &alloc); |
2760 | mhi= multadd(mhi, 10, 0, &alloc); |
2761 | } |
2762 | } |
2763 | } |
2764 | else |
2765 | for (i= 1;; i++) |
2766 | { |
2767 | *s++= dig= quorem(b,S) + '0'; |
2768 | if (!b->p.x[0] && b->wds <= 1) |
2769 | { |
2770 | goto ret; |
2771 | } |
2772 | if (i >= ilim) |
2773 | break; |
2774 | b= multadd(b, 10, 0, &alloc); |
2775 | } |
2776 | |
2777 | /* Round off last digit */ |
2778 | |
2779 | #ifdef Honor_FLT_ROUNDS |
2780 | switch (rounding) { |
2781 | case 0: goto trimzeros; |
2782 | case 2: goto roundoff; |
2783 | } |
2784 | #endif |
2785 | b= lshift(b, 1, &alloc); |
2786 | j= cmp(b, S); |
2787 | if (j > 0 || (j == 0 && dig & 1)) |
2788 | { |
2789 | roundoff: |
2790 | while (*--s == '9') |
2791 | if (s == s0) |
2792 | { |
2793 | k++; |
2794 | *s++= '1'; |
2795 | goto ret; |
2796 | } |
2797 | ++*s++; |
2798 | } |
2799 | else |
2800 | { |
2801 | #ifdef Honor_FLT_ROUNDS |
2802 | trimzeros: |
2803 | #endif |
2804 | while (*--s == '0'); |
2805 | s++; |
2806 | } |
2807 | ret: |
2808 | Bfree(S, &alloc); |
2809 | if (mhi) |
2810 | { |
2811 | if (mlo && mlo != mhi) |
2812 | Bfree(mlo, &alloc); |
2813 | Bfree(mhi, &alloc); |
2814 | } |
2815 | ret1: |
2816 | Bfree(b, &alloc); |
2817 | *s= 0; |
2818 | *decpt= k + 1; |
2819 | if (rve) |
2820 | *rve= s; |
2821 | return s0; |
2822 | } |
2823 | |