Commit | Line | Data |
---|---|---|
98994639 HS |
1 | /* numeric.c |
2 | * | |
3 | * Copyright (c) 2001, Larry Wall | |
4 | * | |
5 | * You may distribute under the terms of either the GNU General Public | |
6 | * License or the Artistic License, as specified in the README file. | |
7 | * | |
8 | */ | |
9 | ||
10 | /* | |
11 | * "That only makes eleven (plus one mislaid) and not fourteen, unless | |
12 | * wizards count differently to other people." | |
13 | */ | |
14 | ||
15 | #include "EXTERN.h" | |
16 | #define PERL_IN_NUMERIC_C | |
17 | #include "perl.h" | |
18 | ||
19 | U32 | |
20 | Perl_cast_ulong(pTHX_ NV f) | |
21 | { | |
22 | if (f < 0.0) | |
23 | return f < I32_MIN ? (U32) I32_MIN : (U32)(I32) f; | |
24 | if (f < U32_MAX_P1) { | |
25 | #if CASTFLAGS & 2 | |
26 | if (f < U32_MAX_P1_HALF) | |
27 | return (U32) f; | |
28 | f -= U32_MAX_P1_HALF; | |
29 | return ((U32) f) | (1 + U32_MAX >> 1); | |
30 | #else | |
31 | return (U32) f; | |
32 | #endif | |
33 | } | |
34 | return f > 0 ? U32_MAX : 0 /* NaN */; | |
35 | } | |
36 | ||
37 | I32 | |
38 | Perl_cast_i32(pTHX_ NV f) | |
39 | { | |
40 | if (f < I32_MAX_P1) | |
41 | return f < I32_MIN ? I32_MIN : (I32) f; | |
42 | if (f < U32_MAX_P1) { | |
43 | #if CASTFLAGS & 2 | |
44 | if (f < U32_MAX_P1_HALF) | |
45 | return (I32)(U32) f; | |
46 | f -= U32_MAX_P1_HALF; | |
47 | return (I32)(((U32) f) | (1 + U32_MAX >> 1)); | |
48 | #else | |
49 | return (I32)(U32) f; | |
50 | #endif | |
51 | } | |
52 | return f > 0 ? (I32)U32_MAX : 0 /* NaN */; | |
53 | } | |
54 | ||
55 | IV | |
56 | Perl_cast_iv(pTHX_ NV f) | |
57 | { | |
58 | if (f < IV_MAX_P1) | |
59 | return f < IV_MIN ? IV_MIN : (IV) f; | |
60 | if (f < UV_MAX_P1) { | |
61 | #if CASTFLAGS & 2 | |
62 | /* For future flexibility allowing for sizeof(UV) >= sizeof(IV) */ | |
63 | if (f < UV_MAX_P1_HALF) | |
64 | return (IV)(UV) f; | |
65 | f -= UV_MAX_P1_HALF; | |
66 | return (IV)(((UV) f) | (1 + UV_MAX >> 1)); | |
67 | #else | |
68 | return (IV)(UV) f; | |
69 | #endif | |
70 | } | |
71 | return f > 0 ? (IV)UV_MAX : 0 /* NaN */; | |
72 | } | |
73 | ||
74 | UV | |
75 | Perl_cast_uv(pTHX_ NV f) | |
76 | { | |
77 | if (f < 0.0) | |
78 | return f < IV_MIN ? (UV) IV_MIN : (UV)(IV) f; | |
79 | if (f < UV_MAX_P1) { | |
80 | #if CASTFLAGS & 2 | |
81 | if (f < UV_MAX_P1_HALF) | |
82 | return (UV) f; | |
83 | f -= UV_MAX_P1_HALF; | |
84 | return ((UV) f) | (1 + UV_MAX >> 1); | |
85 | #else | |
86 | return (UV) f; | |
87 | #endif | |
88 | } | |
89 | return f > 0 ? UV_MAX : 0 /* NaN */; | |
90 | } | |
91 | ||
92 | #if defined(HUGE_VAL) || (defined(USE_LONG_DOUBLE) && defined(HUGE_VALL)) | |
93 | /* | |
94 | * This hack is to force load of "huge" support from libm.a | |
95 | * So it is in perl for (say) POSIX to use. | |
96 | * Needed for SunOS with Sun's 'acc' for example. | |
97 | */ | |
98 | NV | |
99 | Perl_huge(void) | |
100 | { | |
101 | # if defined(USE_LONG_DOUBLE) && defined(HUGE_VALL) | |
102 | return HUGE_VALL; | |
103 | # endif | |
104 | return HUGE_VAL; | |
105 | } | |
106 | #endif | |
107 | ||
108 | NV | |
109 | Perl_scan_bin(pTHX_ char *start, STRLEN len, STRLEN *retlen) | |
110 | { | |
111 | register char *s = start; | |
112 | register NV rnv = 0.0; | |
113 | register UV ruv = 0; | |
114 | register bool seenb = FALSE; | |
115 | register bool overflowed = FALSE; | |
116 | ||
117 | for (; len-- && *s; s++) { | |
118 | if (!(*s == '0' || *s == '1')) { | |
119 | if (*s == '_' && len && *retlen | |
120 | && (s[1] == '0' || s[1] == '1')) | |
121 | { | |
122 | --len; | |
123 | ++s; | |
124 | } | |
125 | else if (seenb == FALSE && *s == 'b' && ruv == 0) { | |
126 | /* Disallow 0bbb0b0bbb... */ | |
127 | seenb = TRUE; | |
128 | continue; | |
129 | } | |
130 | else { | |
131 | if (ckWARN(WARN_DIGIT)) | |
132 | Perl_warner(aTHX_ WARN_DIGIT, | |
133 | "Illegal binary digit '%c' ignored", *s); | |
134 | break; | |
135 | } | |
136 | } | |
137 | if (!overflowed) { | |
138 | register UV xuv = ruv << 1; | |
139 | ||
140 | if ((xuv >> 1) != ruv) { | |
141 | overflowed = TRUE; | |
142 | rnv = (NV) ruv; | |
143 | if (ckWARN_d(WARN_OVERFLOW)) | |
144 | Perl_warner(aTHX_ WARN_OVERFLOW, | |
145 | "Integer overflow in binary number"); | |
146 | } | |
147 | else | |
148 | ruv = xuv | (*s - '0'); | |
149 | } | |
150 | if (overflowed) { | |
151 | rnv *= 2; | |
152 | /* If an NV has not enough bits in its mantissa to | |
153 | * represent an UV this summing of small low-order numbers | |
154 | * is a waste of time (because the NV cannot preserve | |
155 | * the low-order bits anyway): we could just remember when | |
156 | * did we overflow and in the end just multiply rnv by the | |
157 | * right amount. */ | |
158 | rnv += (*s - '0'); | |
159 | } | |
160 | } | |
161 | if (!overflowed) | |
162 | rnv = (NV) ruv; | |
163 | if ( ( overflowed && rnv > 4294967295.0) | |
164 | #if UVSIZE > 4 | |
165 | || (!overflowed && ruv > 0xffffffff ) | |
166 | #endif | |
167 | ) { | |
168 | if (ckWARN(WARN_PORTABLE)) | |
169 | Perl_warner(aTHX_ WARN_PORTABLE, | |
170 | "Binary number > 0b11111111111111111111111111111111 non-portable"); | |
171 | } | |
172 | *retlen = s - start; | |
173 | return rnv; | |
174 | } | |
175 | ||
176 | NV | |
177 | Perl_scan_oct(pTHX_ char *start, STRLEN len, STRLEN *retlen) | |
178 | { | |
179 | register char *s = start; | |
180 | register NV rnv = 0.0; | |
181 | register UV ruv = 0; | |
182 | register bool overflowed = FALSE; | |
183 | ||
184 | for (; len-- && *s; s++) { | |
185 | if (!(*s >= '0' && *s <= '7')) { | |
186 | if (*s == '_' && len && *retlen | |
187 | && (s[1] >= '0' && s[1] <= '7')) | |
188 | { | |
189 | --len; | |
190 | ++s; | |
191 | } | |
192 | else { | |
193 | /* Allow \octal to work the DWIM way (that is, stop scanning | |
194 | * as soon as non-octal characters are seen, complain only iff | |
195 | * someone seems to want to use the digits eight and nine). */ | |
196 | if (*s == '8' || *s == '9') { | |
197 | if (ckWARN(WARN_DIGIT)) | |
198 | Perl_warner(aTHX_ WARN_DIGIT, | |
199 | "Illegal octal digit '%c' ignored", *s); | |
200 | } | |
201 | break; | |
202 | } | |
203 | } | |
204 | if (!overflowed) { | |
205 | register UV xuv = ruv << 3; | |
206 | ||
207 | if ((xuv >> 3) != ruv) { | |
208 | overflowed = TRUE; | |
209 | rnv = (NV) ruv; | |
210 | if (ckWARN_d(WARN_OVERFLOW)) | |
211 | Perl_warner(aTHX_ WARN_OVERFLOW, | |
212 | "Integer overflow in octal number"); | |
213 | } | |
214 | else | |
215 | ruv = xuv | (*s - '0'); | |
216 | } | |
217 | if (overflowed) { | |
218 | rnv *= 8.0; | |
219 | /* If an NV has not enough bits in its mantissa to | |
220 | * represent an UV this summing of small low-order numbers | |
221 | * is a waste of time (because the NV cannot preserve | |
222 | * the low-order bits anyway): we could just remember when | |
223 | * did we overflow and in the end just multiply rnv by the | |
224 | * right amount of 8-tuples. */ | |
225 | rnv += (NV)(*s - '0'); | |
226 | } | |
227 | } | |
228 | if (!overflowed) | |
229 | rnv = (NV) ruv; | |
230 | if ( ( overflowed && rnv > 4294967295.0) | |
231 | #if UVSIZE > 4 | |
232 | || (!overflowed && ruv > 0xffffffff ) | |
233 | #endif | |
234 | ) { | |
235 | if (ckWARN(WARN_PORTABLE)) | |
236 | Perl_warner(aTHX_ WARN_PORTABLE, | |
237 | "Octal number > 037777777777 non-portable"); | |
238 | } | |
239 | *retlen = s - start; | |
240 | return rnv; | |
241 | } | |
242 | ||
243 | NV | |
244 | Perl_scan_hex(pTHX_ char *start, STRLEN len, STRLEN *retlen) | |
245 | { | |
246 | register char *s = start; | |
247 | register NV rnv = 0.0; | |
248 | register UV ruv = 0; | |
249 | register bool overflowed = FALSE; | |
250 | char *hexdigit; | |
251 | ||
252 | if (len > 2) { | |
253 | if (s[0] == 'x') { | |
254 | s++; | |
255 | len--; | |
256 | } | |
257 | else if (len > 3 && s[0] == '0' && s[1] == 'x') { | |
258 | s+=2; | |
259 | len-=2; | |
260 | } | |
261 | } | |
262 | ||
263 | for (; len-- && *s; s++) { | |
264 | hexdigit = strchr((char *) PL_hexdigit, *s); | |
265 | if (!hexdigit) { | |
266 | if (*s == '_' && len && *retlen && s[1] | |
267 | && (hexdigit = strchr((char *) PL_hexdigit, s[1]))) | |
268 | { | |
269 | --len; | |
270 | ++s; | |
271 | } | |
272 | else { | |
273 | if (ckWARN(WARN_DIGIT)) | |
274 | Perl_warner(aTHX_ WARN_DIGIT, | |
275 | "Illegal hexadecimal digit '%c' ignored", *s); | |
276 | break; | |
277 | } | |
278 | } | |
279 | if (!overflowed) { | |
280 | register UV xuv = ruv << 4; | |
281 | ||
282 | if ((xuv >> 4) != ruv) { | |
283 | overflowed = TRUE; | |
284 | rnv = (NV) ruv; | |
285 | if (ckWARN_d(WARN_OVERFLOW)) | |
286 | Perl_warner(aTHX_ WARN_OVERFLOW, | |
287 | "Integer overflow in hexadecimal number"); | |
288 | } | |
289 | else | |
290 | ruv = xuv | ((hexdigit - PL_hexdigit) & 15); | |
291 | } | |
292 | if (overflowed) { | |
293 | rnv *= 16.0; | |
294 | /* If an NV has not enough bits in its mantissa to | |
295 | * represent an UV this summing of small low-order numbers | |
296 | * is a waste of time (because the NV cannot preserve | |
297 | * the low-order bits anyway): we could just remember when | |
298 | * did we overflow and in the end just multiply rnv by the | |
299 | * right amount of 16-tuples. */ | |
300 | rnv += (NV)((hexdigit - PL_hexdigit) & 15); | |
301 | } | |
302 | } | |
303 | if (!overflowed) | |
304 | rnv = (NV) ruv; | |
305 | if ( ( overflowed && rnv > 4294967295.0) | |
306 | #if UVSIZE > 4 | |
307 | || (!overflowed && ruv > 0xffffffff ) | |
308 | #endif | |
309 | ) { | |
310 | if (ckWARN(WARN_PORTABLE)) | |
311 | Perl_warner(aTHX_ WARN_PORTABLE, | |
312 | "Hexadecimal number > 0xffffffff non-portable"); | |
313 | } | |
314 | *retlen = s - start; | |
315 | return rnv; | |
316 | } | |
317 | ||
318 | /* | |
319 | =for apidoc grok_numeric_radix | |
320 | ||
321 | Scan and skip for a numeric decimal separator (radix). | |
322 | ||
323 | =cut | |
324 | */ | |
325 | bool | |
326 | Perl_grok_numeric_radix(pTHX_ const char **sp, const char *send) | |
327 | { | |
328 | #ifdef USE_LOCALE_NUMERIC | |
329 | if (PL_numeric_radix_sv && IN_LOCALE) { | |
330 | STRLEN len; | |
331 | char* radix = SvPV(PL_numeric_radix_sv, len); | |
332 | if (*sp + len <= send && memEQ(*sp, radix, len)) { | |
333 | *sp += len; | |
334 | return TRUE; | |
335 | } | |
336 | } | |
337 | /* always try "." if numeric radix didn't match because | |
338 | * we may have data from different locales mixed */ | |
339 | #endif | |
340 | if (*sp < send && **sp == '.') { | |
341 | ++*sp; | |
342 | return TRUE; | |
343 | } | |
344 | return FALSE; | |
345 | } | |
346 | ||
347 | /* | |
348 | =for apidoc grok_number | |
349 | ||
350 | Recognise (or not) a number. The type of the number is returned | |
351 | (0 if unrecognised), otherwise it is a bit-ORed combination of | |
352 | IS_NUMBER_IN_UV, IS_NUMBER_GREATER_THAN_UV_MAX, IS_NUMBER_NOT_INT, | |
aa8b85de | 353 | IS_NUMBER_NEG, IS_NUMBER_INFINITY, IS_NUMBER_NAN (defined in perl.h). |
60939fb8 NC |
354 | |
355 | If the value of the number can fit an in UV, it is returned in the *valuep | |
356 | IS_NUMBER_IN_UV will be set to indicate that *valuep is valid, IS_NUMBER_IN_UV | |
357 | will never be set unless *valuep is valid, but *valuep may have been assigned | |
358 | to during processing even though IS_NUMBER_IN_UV is not set on return. | |
359 | If valuep is NULL, IS_NUMBER_IN_UV will be set for the same cases as when | |
360 | valuep is non-NULL, but no actual assignment (or SEGV) will occur. | |
361 | ||
362 | IS_NUMBER_NOT_INT will be set with IS_NUMBER_IN_UV if trailing decimals were | |
363 | seen (in which case *valuep gives the true value truncated to an integer), and | |
364 | IS_NUMBER_NEG if the number is negative (in which case *valuep holds the | |
365 | absolute value). IS_NUMBER_IN_UV is not set if e notation was used or the | |
366 | number is larger than a UV. | |
98994639 HS |
367 | |
368 | =cut | |
369 | */ | |
370 | int | |
371 | Perl_grok_number(pTHX_ const char *pv, STRLEN len, UV *valuep) | |
372 | { | |
60939fb8 NC |
373 | const char *s = pv; |
374 | const char *send = pv + len; | |
375 | const UV max_div_10 = UV_MAX / 10; | |
376 | const char max_mod_10 = UV_MAX % 10; | |
377 | int numtype = 0; | |
378 | int sawinf = 0; | |
aa8b85de | 379 | int sawnan = 0; |
60939fb8 NC |
380 | |
381 | while (s < send && isSPACE(*s)) | |
382 | s++; | |
383 | if (s == send) { | |
384 | return 0; | |
385 | } else if (*s == '-') { | |
386 | s++; | |
387 | numtype = IS_NUMBER_NEG; | |
388 | } | |
389 | else if (*s == '+') | |
390 | s++; | |
391 | ||
392 | if (s == send) | |
393 | return 0; | |
394 | ||
395 | /* next must be digit or the radix separator or beginning of infinity */ | |
396 | if (isDIGIT(*s)) { | |
397 | /* UVs are at least 32 bits, so the first 9 decimal digits cannot | |
398 | overflow. */ | |
399 | UV value = *s - '0'; | |
400 | /* This construction seems to be more optimiser friendly. | |
401 | (without it gcc does the isDIGIT test and the *s - '0' separately) | |
402 | With it gcc on arm is managing 6 instructions (6 cycles) per digit. | |
403 | In theory the optimiser could deduce how far to unroll the loop | |
404 | before checking for overflow. */ | |
58bb9ec3 NC |
405 | if (++s < send) { |
406 | int digit = *s - '0'; | |
60939fb8 NC |
407 | if (digit >= 0 && digit <= 9) { |
408 | value = value * 10 + digit; | |
58bb9ec3 NC |
409 | if (++s < send) { |
410 | digit = *s - '0'; | |
60939fb8 NC |
411 | if (digit >= 0 && digit <= 9) { |
412 | value = value * 10 + digit; | |
58bb9ec3 NC |
413 | if (++s < send) { |
414 | digit = *s - '0'; | |
60939fb8 NC |
415 | if (digit >= 0 && digit <= 9) { |
416 | value = value * 10 + digit; | |
58bb9ec3 NC |
417 | if (++s < send) { |
418 | digit = *s - '0'; | |
60939fb8 NC |
419 | if (digit >= 0 && digit <= 9) { |
420 | value = value * 10 + digit; | |
58bb9ec3 NC |
421 | if (++s < send) { |
422 | digit = *s - '0'; | |
60939fb8 NC |
423 | if (digit >= 0 && digit <= 9) { |
424 | value = value * 10 + digit; | |
58bb9ec3 NC |
425 | if (++s < send) { |
426 | digit = *s - '0'; | |
60939fb8 NC |
427 | if (digit >= 0 && digit <= 9) { |
428 | value = value * 10 + digit; | |
58bb9ec3 NC |
429 | if (++s < send) { |
430 | digit = *s - '0'; | |
60939fb8 NC |
431 | if (digit >= 0 && digit <= 9) { |
432 | value = value * 10 + digit; | |
58bb9ec3 NC |
433 | if (++s < send) { |
434 | digit = *s - '0'; | |
60939fb8 NC |
435 | if (digit >= 0 && digit <= 9) { |
436 | value = value * 10 + digit; | |
58bb9ec3 | 437 | if (++s < send) { |
60939fb8 NC |
438 | /* Now got 9 digits, so need to check |
439 | each time for overflow. */ | |
58bb9ec3 | 440 | digit = *s - '0'; |
60939fb8 NC |
441 | while (digit >= 0 && digit <= 9 |
442 | && (value < max_div_10 | |
443 | || (value == max_div_10 | |
444 | && digit <= max_mod_10))) { | |
445 | value = value * 10 + digit; | |
58bb9ec3 NC |
446 | if (++s < send) |
447 | digit = *s - '0'; | |
60939fb8 NC |
448 | else |
449 | break; | |
450 | } | |
451 | if (digit >= 0 && digit <= 9 | |
51bd16da | 452 | && (s < send)) { |
60939fb8 NC |
453 | /* value overflowed. |
454 | skip the remaining digits, don't | |
455 | worry about setting *valuep. */ | |
456 | do { | |
457 | s++; | |
458 | } while (s < send && isDIGIT(*s)); | |
459 | numtype |= | |
460 | IS_NUMBER_GREATER_THAN_UV_MAX; | |
461 | goto skip_value; | |
462 | } | |
463 | } | |
464 | } | |
98994639 | 465 | } |
60939fb8 NC |
466 | } |
467 | } | |
468 | } | |
469 | } | |
470 | } | |
471 | } | |
472 | } | |
473 | } | |
474 | } | |
475 | } | |
476 | } | |
98994639 | 477 | } |
60939fb8 | 478 | } |
98994639 | 479 | } |
60939fb8 NC |
480 | numtype |= IS_NUMBER_IN_UV; |
481 | if (valuep) | |
482 | *valuep = value; | |
483 | ||
484 | skip_value: | |
485 | if (GROK_NUMERIC_RADIX(&s, send)) { | |
486 | numtype |= IS_NUMBER_NOT_INT; | |
487 | while (s < send && isDIGIT(*s)) /* optional digits after the radix */ | |
488 | s++; | |
98994639 | 489 | } |
60939fb8 NC |
490 | } |
491 | else if (GROK_NUMERIC_RADIX(&s, send)) { | |
492 | numtype |= IS_NUMBER_NOT_INT | IS_NUMBER_IN_UV; /* valuep assigned below */ | |
493 | /* no digits before the radix means we need digits after it */ | |
494 | if (s < send && isDIGIT(*s)) { | |
495 | do { | |
496 | s++; | |
497 | } while (s < send && isDIGIT(*s)); | |
498 | if (valuep) { | |
499 | /* integer approximation is valid - it's 0. */ | |
500 | *valuep = 0; | |
501 | } | |
98994639 | 502 | } |
60939fb8 NC |
503 | else |
504 | return 0; | |
505 | } else if (*s == 'I' || *s == 'i') { | |
506 | s++; if (s == send || (*s != 'N' && *s != 'n')) return 0; | |
507 | s++; if (s == send || (*s != 'F' && *s != 'f')) return 0; | |
508 | s++; if (s < send && (*s == 'I' || *s == 'i')) { | |
509 | s++; if (s == send || (*s != 'N' && *s != 'n')) return 0; | |
510 | s++; if (s == send || (*s != 'I' && *s != 'i')) return 0; | |
511 | s++; if (s == send || (*s != 'T' && *s != 't')) return 0; | |
512 | s++; if (s == send || (*s != 'Y' && *s != 'y')) return 0; | |
513 | s++; | |
98994639 | 514 | } |
60939fb8 | 515 | sawinf = 1; |
aa8b85de JH |
516 | } else if (*s == 'N' || *s == 'n') { |
517 | /* XXX TODO: There are signaling NaNs and quiet NaNs. */ | |
518 | s++; if (s == send || (*s != 'A' && *s != 'a')) return 0; | |
519 | s++; if (s == send || (*s != 'N' && *s != 'n')) return 0; | |
520 | s++; | |
521 | sawnan = 1; | |
522 | } else | |
98994639 | 523 | return 0; |
60939fb8 NC |
524 | |
525 | if (sawinf) { | |
526 | numtype &= IS_NUMBER_NEG; /* Keep track of sign */ | |
527 | numtype |= IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT; | |
aa8b85de JH |
528 | } else if (sawnan) { |
529 | numtype &= IS_NUMBER_NEG; /* Keep track of sign */ | |
530 | numtype |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT; | |
60939fb8 NC |
531 | } else if (s < send) { |
532 | /* we can have an optional exponent part */ | |
533 | if (*s == 'e' || *s == 'E') { | |
534 | /* The only flag we keep is sign. Blow away any "it's UV" */ | |
535 | numtype &= IS_NUMBER_NEG; | |
536 | numtype |= IS_NUMBER_NOT_INT; | |
537 | s++; | |
538 | if (s < send && (*s == '-' || *s == '+')) | |
539 | s++; | |
540 | if (s < send && isDIGIT(*s)) { | |
541 | do { | |
542 | s++; | |
543 | } while (s < send && isDIGIT(*s)); | |
544 | } | |
545 | else | |
546 | return 0; | |
547 | } | |
548 | } | |
549 | while (s < send && isSPACE(*s)) | |
550 | s++; | |
551 | if (s >= send) | |
aa8b85de | 552 | return numtype; |
60939fb8 NC |
553 | if (len == 10 && memEQ(pv, "0 but true", 10)) { |
554 | if (valuep) | |
555 | *valuep = 0; | |
556 | return IS_NUMBER_IN_UV; | |
557 | } | |
558 | return 0; | |
98994639 HS |
559 | } |
560 | ||
561 | NV | |
562 | S_mulexp10(NV value, I32 exponent) | |
563 | { | |
564 | NV result = 1.0; | |
565 | NV power = 10.0; | |
566 | bool negative = 0; | |
567 | I32 bit; | |
568 | ||
569 | if (exponent == 0) | |
570 | return value; | |
571 | else if (exponent < 0) { | |
572 | negative = 1; | |
573 | exponent = -exponent; | |
574 | } | |
87032ba1 | 575 | |
24866caa | 576 | /* On OpenVMS VAX we by default use the D_FLOAT double format, |
67597c89 | 577 | * and that format does not have *easy* capabilities [1] for |
24866caa CB |
578 | * overflowing doubles 'silently' as IEEE fp does. We also need |
579 | * to support G_FLOAT on both VAX and Alpha, and though the exponent | |
580 | * range is much larger than D_FLOAT it still doesn't do silent | |
581 | * overflow. Therefore we need to detect early whether we would | |
582 | * overflow (this is the behaviour of the native string-to-float | |
583 | * conversion routines, and therefore of native applications, too). | |
67597c89 | 584 | * |
24866caa CB |
585 | * [1] Trying to establish a condition handler to trap floating point |
586 | * exceptions is not a good idea. */ | |
587 | #if defined(VMS) && !defined(__IEEE_FP) && defined(NV_MAX_10_EXP) | |
67597c89 | 588 | if (!negative && |
24866caa | 589 | (log10(value) + exponent) >= (NV_MAX_10_EXP)) |
67597c89 | 590 | return NV_MAX; |
67597c89 | 591 | #endif |
87032ba1 JH |
592 | |
593 | /* In UNICOS and in certain Cray models (such as T90) there is no | |
594 | * IEEE fp, and no way at all from C to catch fp overflows gracefully. | |
595 | * There is something you can do if you are willing to use some | |
596 | * inline assembler: the instruction is called DFI-- but that will | |
597 | * disable *all* floating point interrupts, a little bit too large | |
598 | * a hammer. Therefore we need to catch potential overflows before | |
599 | * it's too late. */ | |
600 | #if defined(_UNICOS) && defined(NV_MAX_10_EXP) | |
601 | if (!negative && | |
602 | (log10(value) + exponent) >= NV_MAX_10_EXP) | |
603 | return NV_MAX; | |
604 | #endif | |
605 | ||
98994639 HS |
606 | for (bit = 1; exponent; bit <<= 1) { |
607 | if (exponent & bit) { | |
608 | exponent ^= bit; | |
609 | result *= power; | |
610 | } | |
7014c407 | 611 | /* Floating point exceptions are supposed to be turned off. */ |
98994639 HS |
612 | power *= power; |
613 | } | |
614 | return negative ? value / result : value * result; | |
615 | } | |
616 | ||
617 | NV | |
618 | Perl_my_atof(pTHX_ const char* s) | |
619 | { | |
620 | NV x = 0.0; | |
621 | #ifdef USE_LOCALE_NUMERIC | |
622 | if (PL_numeric_local && IN_LOCALE) { | |
623 | NV y; | |
624 | ||
625 | /* Scan the number twice; once using locale and once without; | |
626 | * choose the larger result (in absolute value). */ | |
627 | Perl_atof2(aTHX_ s, &x); | |
628 | SET_NUMERIC_STANDARD(); | |
629 | Perl_atof2(aTHX_ s, &y); | |
630 | SET_NUMERIC_LOCAL(); | |
631 | if ((y < 0.0 && y < x) || (y > 0.0 && y > x)) | |
632 | return y; | |
633 | } | |
634 | else | |
635 | Perl_atof2(aTHX_ s, &x); | |
636 | #else | |
637 | Perl_atof2(aTHX_ s, &x); | |
638 | #endif | |
639 | return x; | |
640 | } | |
641 | ||
642 | char* | |
643 | Perl_my_atof2(pTHX_ const char* orig, NV* value) | |
644 | { | |
645 | NV result = 0.0; | |
646 | bool negative = 0; | |
647 | char* s = (char*)orig; | |
648 | char* send = s + strlen(orig) - 1; | |
649 | bool seendigit = 0; | |
650 | I32 expextra = 0; | |
651 | I32 exponent = 0; | |
652 | I32 i; | |
653 | /* this is arbitrary */ | |
654 | #define PARTLIM 6 | |
655 | /* we want the largest integers we can usefully use */ | |
656 | #if defined(HAS_QUAD) && defined(USE_64_BIT_INT) | |
657 | # define PARTSIZE ((int)TYPE_DIGITS(U64)-1) | |
658 | U64 part[PARTLIM]; | |
659 | #else | |
660 | # define PARTSIZE ((int)TYPE_DIGITS(U32)-1) | |
661 | U32 part[PARTLIM]; | |
662 | #endif | |
663 | I32 ipart = 0; /* index into part[] */ | |
664 | I32 offcount; /* number of digits in least significant part */ | |
665 | ||
96a05aee HS |
666 | /* leading whitespace */ |
667 | while (isSPACE(*s)) | |
668 | ++s; | |
669 | ||
98994639 HS |
670 | /* sign */ |
671 | switch (*s) { | |
672 | case '-': | |
673 | negative = 1; | |
674 | /* fall through */ | |
675 | case '+': | |
676 | ++s; | |
677 | } | |
678 | ||
679 | part[0] = offcount = 0; | |
680 | if (isDIGIT(*s)) { | |
681 | seendigit = 1; /* get this over with */ | |
682 | ||
683 | /* skip leading zeros */ | |
684 | while (*s == '0') | |
685 | ++s; | |
686 | } | |
687 | ||
688 | /* integer digits */ | |
689 | while (isDIGIT(*s)) { | |
690 | if (++offcount > PARTSIZE) { | |
691 | if (++ipart < PARTLIM) { | |
692 | part[ipart] = 0; | |
693 | offcount = 1; /* ++0 */ | |
694 | } | |
695 | else { | |
696 | /* limits of precision reached */ | |
697 | --ipart; | |
698 | --offcount; | |
699 | if (*s >= '5') | |
700 | ++part[ipart]; | |
701 | while (isDIGIT(*s)) { | |
702 | ++expextra; | |
703 | ++s; | |
704 | } | |
705 | /* warn of loss of precision? */ | |
706 | break; | |
707 | } | |
708 | } | |
709 | part[ipart] = part[ipart] * 10 + (*s++ - '0'); | |
710 | } | |
711 | ||
712 | /* decimal point */ | |
713 | if (GROK_NUMERIC_RADIX((const char **)&s, send)) { | |
714 | if (isDIGIT(*s)) | |
715 | seendigit = 1; /* get this over with */ | |
716 | ||
717 | /* decimal digits */ | |
718 | while (isDIGIT(*s)) { | |
719 | if (++offcount > PARTSIZE) { | |
720 | if (++ipart < PARTLIM) { | |
721 | part[ipart] = 0; | |
722 | offcount = 1; /* ++0 */ | |
723 | } | |
724 | else { | |
725 | /* limits of precision reached */ | |
726 | --ipart; | |
727 | --offcount; | |
728 | if (*s >= '5') | |
729 | ++part[ipart]; | |
730 | while (isDIGIT(*s)) | |
731 | ++s; | |
732 | /* warn of loss of precision? */ | |
733 | break; | |
734 | } | |
735 | } | |
736 | --expextra; | |
737 | part[ipart] = part[ipart] * 10 + (*s++ - '0'); | |
738 | } | |
739 | } | |
740 | ||
741 | /* combine components of mantissa */ | |
742 | for (i = 0; i <= ipart; ++i) | |
743 | result += S_mulexp10((NV)part[ipart - i], | |
744 | i ? offcount + (i - 1) * PARTSIZE : 0); | |
745 | ||
746 | if (seendigit && (*s == 'e' || *s == 'E')) { | |
747 | bool expnegative = 0; | |
748 | ||
749 | ++s; | |
750 | switch (*s) { | |
751 | case '-': | |
752 | expnegative = 1; | |
753 | /* fall through */ | |
754 | case '+': | |
755 | ++s; | |
756 | } | |
757 | while (isDIGIT(*s)) | |
758 | exponent = exponent * 10 + (*s++ - '0'); | |
759 | if (expnegative) | |
760 | exponent = -exponent; | |
761 | } | |
762 | ||
763 | /* now apply the exponent */ | |
764 | exponent += expextra; | |
765 | result = S_mulexp10(result, exponent); | |
766 | ||
767 | /* now apply the sign */ | |
768 | if (negative) | |
769 | result = -result; | |
770 | *value = result; | |
771 | return s; | |
772 | } | |
773 |