This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
perldelta: Use '#' for ticket number
[perl5.git] / numeric.c
... / ...
CommitLineData
1/* numeric.c
2 *
3 * Copyright (C) 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001,
4 * 2002, 2003, 2004, 2005, 2006, 2007, 2008 by Larry Wall and others
5 *
6 * You may distribute under the terms of either the GNU General Public
7 * License or the Artistic License, as specified in the README file.
8 *
9 */
10
11/*
12 * "That only makes eleven (plus one mislaid) and not fourteen,
13 * unless wizards count differently to other people." --Beorn
14 *
15 * [p.115 of _The Hobbit_: "Queer Lodgings"]
16 */
17
18/*
19=head1 Numeric functions
20
21=cut
22
23This file contains all the stuff needed by perl for manipulating numeric
24values, including such things as replacements for the OS's atof() function
25
26*/
27
28#include "EXTERN.h"
29#define PERL_IN_NUMERIC_C
30#include "perl.h"
31
32U32
33Perl_cast_ulong(NV f)
34{
35 if (f < 0.0)
36 return f < I32_MIN ? (U32) I32_MIN : (U32)(I32) f;
37 if (f < U32_MAX_P1) {
38#if CASTFLAGS & 2
39 if (f < U32_MAX_P1_HALF)
40 return (U32) f;
41 f -= U32_MAX_P1_HALF;
42 return ((U32) f) | (1 + (U32_MAX >> 1));
43#else
44 return (U32) f;
45#endif
46 }
47 return f > 0 ? U32_MAX : 0 /* NaN */;
48}
49
50I32
51Perl_cast_i32(NV f)
52{
53 if (f < I32_MAX_P1)
54 return f < I32_MIN ? I32_MIN : (I32) f;
55 if (f < U32_MAX_P1) {
56#if CASTFLAGS & 2
57 if (f < U32_MAX_P1_HALF)
58 return (I32)(U32) f;
59 f -= U32_MAX_P1_HALF;
60 return (I32)(((U32) f) | (1 + (U32_MAX >> 1)));
61#else
62 return (I32)(U32) f;
63#endif
64 }
65 return f > 0 ? (I32)U32_MAX : 0 /* NaN */;
66}
67
68IV
69Perl_cast_iv(NV f)
70{
71 if (f < IV_MAX_P1)
72 return f < IV_MIN ? IV_MIN : (IV) f;
73 if (f < UV_MAX_P1) {
74#if CASTFLAGS & 2
75 /* For future flexibility allowing for sizeof(UV) >= sizeof(IV) */
76 if (f < UV_MAX_P1_HALF)
77 return (IV)(UV) f;
78 f -= UV_MAX_P1_HALF;
79 return (IV)(((UV) f) | (1 + (UV_MAX >> 1)));
80#else
81 return (IV)(UV) f;
82#endif
83 }
84 return f > 0 ? (IV)UV_MAX : 0 /* NaN */;
85}
86
87UV
88Perl_cast_uv(NV f)
89{
90 if (f < 0.0)
91 return f < IV_MIN ? (UV) IV_MIN : (UV)(IV) f;
92 if (f < UV_MAX_P1) {
93#if CASTFLAGS & 2
94 if (f < UV_MAX_P1_HALF)
95 return (UV) f;
96 f -= UV_MAX_P1_HALF;
97 return ((UV) f) | (1 + (UV_MAX >> 1));
98#else
99 return (UV) f;
100#endif
101 }
102 return f > 0 ? UV_MAX : 0 /* NaN */;
103}
104
105/*
106=for apidoc grok_bin
107
108converts a string representing a binary number to numeric form.
109
110On entry C<start> and C<*len> give the string to scan, C<*flags> gives
111conversion flags, and C<result> should be C<NULL> or a pointer to an NV.
112The scan stops at the end of the string, or the first invalid character.
113Unless C<PERL_SCAN_SILENT_ILLDIGIT> is set in C<*flags>, encountering an
114invalid character will also trigger a warning.
115On return C<*len> is set to the length of the scanned string,
116and C<*flags> gives output flags.
117
118If the value is <= C<UV_MAX> it is returned as a UV, the output flags are clear,
119and nothing is written to C<*result>. If the value is > C<UV_MAX>, C<grok_bin>
120returns C<UV_MAX>, sets C<PERL_SCAN_GREATER_THAN_UV_MAX> in the output flags,
121and writes the value to C<*result> (or the value is discarded if C<result>
122is NULL).
123
124The binary number may optionally be prefixed with C<"0b"> or C<"b"> unless
125C<PERL_SCAN_DISALLOW_PREFIX> is set in C<*flags> on entry. If
126C<PERL_SCAN_ALLOW_UNDERSCORES> is set in C<*flags> then the binary
127number may use C<"_"> characters to separate digits.
128
129=cut
130
131Not documented yet because experimental is C<PERL_SCAN_SILENT_NON_PORTABLE
132which suppresses any message for non-portable numbers that are still valid
133on this platform.
134 */
135
136UV
137Perl_grok_bin(pTHX_ const char *start, STRLEN *len_p, I32 *flags, NV *result)
138{
139 const char *s = start;
140 STRLEN len = *len_p;
141 UV value = 0;
142 NV value_nv = 0;
143
144 const UV max_div_2 = UV_MAX / 2;
145 const bool allow_underscores = cBOOL(*flags & PERL_SCAN_ALLOW_UNDERSCORES);
146 bool overflowed = FALSE;
147 char bit;
148
149 PERL_ARGS_ASSERT_GROK_BIN;
150
151 if (!(*flags & PERL_SCAN_DISALLOW_PREFIX)) {
152 /* strip off leading b or 0b.
153 for compatibility silently suffer "b" and "0b" as valid binary
154 numbers. */
155 if (len >= 1) {
156 if (isALPHA_FOLD_EQ(s[0], 'b')) {
157 s++;
158 len--;
159 }
160 else if (len >= 2 && s[0] == '0' && (isALPHA_FOLD_EQ(s[1], 'b'))) {
161 s+=2;
162 len-=2;
163 }
164 }
165 }
166
167 for (; len-- && (bit = *s); s++) {
168 if (bit == '0' || bit == '1') {
169 /* Write it in this wonky order with a goto to attempt to get the
170 compiler to make the common case integer-only loop pretty tight.
171 With gcc seems to be much straighter code than old scan_bin. */
172 redo:
173 if (!overflowed) {
174 if (value <= max_div_2) {
175 value = (value << 1) | (bit - '0');
176 continue;
177 }
178 /* Bah. We're just overflowed. */
179 /* diag_listed_as: Integer overflow in %s number */
180 Perl_ck_warner_d(aTHX_ packWARN(WARN_OVERFLOW),
181 "Integer overflow in binary number");
182 overflowed = TRUE;
183 value_nv = (NV) value;
184 }
185 value_nv *= 2.0;
186 /* If an NV has not enough bits in its mantissa to
187 * represent a UV this summing of small low-order numbers
188 * is a waste of time (because the NV cannot preserve
189 * the low-order bits anyway): we could just remember when
190 * did we overflow and in the end just multiply value_nv by the
191 * right amount. */
192 value_nv += (NV)(bit - '0');
193 continue;
194 }
195 if (bit == '_' && len && allow_underscores && (bit = s[1])
196 && (bit == '0' || bit == '1'))
197 {
198 --len;
199 ++s;
200 goto redo;
201 }
202 if (!(*flags & PERL_SCAN_SILENT_ILLDIGIT))
203 Perl_ck_warner(aTHX_ packWARN(WARN_DIGIT),
204 "Illegal binary digit '%c' ignored", *s);
205 break;
206 }
207
208 if ( ( overflowed && value_nv > 4294967295.0)
209#if UVSIZE > 4
210 || (!overflowed && value > 0xffffffff
211 && ! (*flags & PERL_SCAN_SILENT_NON_PORTABLE))
212#endif
213 ) {
214 Perl_ck_warner(aTHX_ packWARN(WARN_PORTABLE),
215 "Binary number > 0b11111111111111111111111111111111 non-portable");
216 }
217 *len_p = s - start;
218 if (!overflowed) {
219 *flags = 0;
220 return value;
221 }
222 *flags = PERL_SCAN_GREATER_THAN_UV_MAX;
223 if (result)
224 *result = value_nv;
225 return UV_MAX;
226}
227
228/*
229=for apidoc grok_hex
230
231converts a string representing a hex number to numeric form.
232
233On entry C<start> and C<*len_p> give the string to scan, C<*flags> gives
234conversion flags, and C<result> should be C<NULL> or a pointer to an NV.
235The scan stops at the end of the string, or the first invalid character.
236Unless C<PERL_SCAN_SILENT_ILLDIGIT> is set in C<*flags>, encountering an
237invalid character will also trigger a warning.
238On return C<*len> is set to the length of the scanned string,
239and C<*flags> gives output flags.
240
241If the value is <= C<UV_MAX> it is returned as a UV, the output flags are clear,
242and nothing is written to C<*result>. If the value is > C<UV_MAX>, C<grok_hex>
243returns C<UV_MAX>, sets C<PERL_SCAN_GREATER_THAN_UV_MAX> in the output flags,
244and writes the value to C<*result> (or the value is discarded if C<result>
245is C<NULL>).
246
247The hex number may optionally be prefixed with C<"0x"> or C<"x"> unless
248C<PERL_SCAN_DISALLOW_PREFIX> is set in C<*flags> on entry. If
249C<PERL_SCAN_ALLOW_UNDERSCORES> is set in C<*flags> then the hex
250number may use C<"_"> characters to separate digits.
251
252=cut
253
254Not documented yet because experimental is C<PERL_SCAN_SILENT_NON_PORTABLE
255which suppresses any message for non-portable numbers, but which are valid
256on this platform.
257 */
258
259UV
260Perl_grok_hex(pTHX_ const char *start, STRLEN *len_p, I32 *flags, NV *result)
261{
262 const char *s = start;
263 STRLEN len = *len_p;
264 UV value = 0;
265 NV value_nv = 0;
266 const UV max_div_16 = UV_MAX / 16;
267 const bool allow_underscores = cBOOL(*flags & PERL_SCAN_ALLOW_UNDERSCORES);
268 bool overflowed = FALSE;
269
270 PERL_ARGS_ASSERT_GROK_HEX;
271
272 if (!(*flags & PERL_SCAN_DISALLOW_PREFIX)) {
273 /* strip off leading x or 0x.
274 for compatibility silently suffer "x" and "0x" as valid hex numbers.
275 */
276 if (len >= 1) {
277 if (isALPHA_FOLD_EQ(s[0], 'x')) {
278 s++;
279 len--;
280 }
281 else if (len >= 2 && s[0] == '0' && (isALPHA_FOLD_EQ(s[1], 'x'))) {
282 s+=2;
283 len-=2;
284 }
285 }
286 }
287
288 for (; len-- && *s; s++) {
289 if (isXDIGIT(*s)) {
290 /* Write it in this wonky order with a goto to attempt to get the
291 compiler to make the common case integer-only loop pretty tight.
292 With gcc seems to be much straighter code than old scan_hex. */
293 redo:
294 if (!overflowed) {
295 if (value <= max_div_16) {
296 value = (value << 4) | XDIGIT_VALUE(*s);
297 continue;
298 }
299 /* Bah. We're just overflowed. */
300 /* diag_listed_as: Integer overflow in %s number */
301 Perl_ck_warner_d(aTHX_ packWARN(WARN_OVERFLOW),
302 "Integer overflow in hexadecimal number");
303 overflowed = TRUE;
304 value_nv = (NV) value;
305 }
306 value_nv *= 16.0;
307 /* If an NV has not enough bits in its mantissa to
308 * represent a UV this summing of small low-order numbers
309 * is a waste of time (because the NV cannot preserve
310 * the low-order bits anyway): we could just remember when
311 * did we overflow and in the end just multiply value_nv by the
312 * right amount of 16-tuples. */
313 value_nv += (NV) XDIGIT_VALUE(*s);
314 continue;
315 }
316 if (*s == '_' && len && allow_underscores && s[1]
317 && isXDIGIT(s[1]))
318 {
319 --len;
320 ++s;
321 goto redo;
322 }
323 if (!(*flags & PERL_SCAN_SILENT_ILLDIGIT))
324 Perl_ck_warner(aTHX_ packWARN(WARN_DIGIT),
325 "Illegal hexadecimal digit '%c' ignored", *s);
326 break;
327 }
328
329 if ( ( overflowed && value_nv > 4294967295.0)
330#if UVSIZE > 4
331 || (!overflowed && value > 0xffffffff
332 && ! (*flags & PERL_SCAN_SILENT_NON_PORTABLE))
333#endif
334 ) {
335 Perl_ck_warner(aTHX_ packWARN(WARN_PORTABLE),
336 "Hexadecimal number > 0xffffffff non-portable");
337 }
338 *len_p = s - start;
339 if (!overflowed) {
340 *flags = 0;
341 return value;
342 }
343 *flags = PERL_SCAN_GREATER_THAN_UV_MAX;
344 if (result)
345 *result = value_nv;
346 return UV_MAX;
347}
348
349/*
350=for apidoc grok_oct
351
352converts a string representing an octal number to numeric form.
353
354On entry C<start> and C<*len> give the string to scan, C<*flags> gives
355conversion flags, and C<result> should be C<NULL> or a pointer to an NV.
356The scan stops at the end of the string, or the first invalid character.
357Unless C<PERL_SCAN_SILENT_ILLDIGIT> is set in C<*flags>, encountering an
3588 or 9 will also trigger a warning.
359On return C<*len> is set to the length of the scanned string,
360and C<*flags> gives output flags.
361
362If the value is <= C<UV_MAX> it is returned as a UV, the output flags are clear,
363and nothing is written to C<*result>. If the value is > C<UV_MAX>, C<grok_oct>
364returns C<UV_MAX>, sets C<PERL_SCAN_GREATER_THAN_UV_MAX> in the output flags,
365and writes the value to C<*result> (or the value is discarded if C<result>
366is C<NULL>).
367
368If C<PERL_SCAN_ALLOW_UNDERSCORES> is set in C<*flags> then the octal
369number may use C<"_"> characters to separate digits.
370
371=cut
372
373Not documented yet because experimental is C<PERL_SCAN_SILENT_NON_PORTABLE>
374which suppresses any message for non-portable numbers, but which are valid
375on this platform.
376 */
377
378UV
379Perl_grok_oct(pTHX_ const char *start, STRLEN *len_p, I32 *flags, NV *result)
380{
381 const char *s = start;
382 STRLEN len = *len_p;
383 UV value = 0;
384 NV value_nv = 0;
385 const UV max_div_8 = UV_MAX / 8;
386 const bool allow_underscores = cBOOL(*flags & PERL_SCAN_ALLOW_UNDERSCORES);
387 bool overflowed = FALSE;
388
389 PERL_ARGS_ASSERT_GROK_OCT;
390
391 for (; len-- && *s; s++) {
392 if (isOCTAL(*s)) {
393 /* Write it in this wonky order with a goto to attempt to get the
394 compiler to make the common case integer-only loop pretty tight.
395 */
396 redo:
397 if (!overflowed) {
398 if (value <= max_div_8) {
399 value = (value << 3) | OCTAL_VALUE(*s);
400 continue;
401 }
402 /* Bah. We're just overflowed. */
403 /* diag_listed_as: Integer overflow in %s number */
404 Perl_ck_warner_d(aTHX_ packWARN(WARN_OVERFLOW),
405 "Integer overflow in octal number");
406 overflowed = TRUE;
407 value_nv = (NV) value;
408 }
409 value_nv *= 8.0;
410 /* If an NV has not enough bits in its mantissa to
411 * represent a UV this summing of small low-order numbers
412 * is a waste of time (because the NV cannot preserve
413 * the low-order bits anyway): we could just remember when
414 * did we overflow and in the end just multiply value_nv by the
415 * right amount of 8-tuples. */
416 value_nv += (NV) OCTAL_VALUE(*s);
417 continue;
418 }
419 if (*s == '_' && len && allow_underscores && isOCTAL(s[1])) {
420 --len;
421 ++s;
422 goto redo;
423 }
424 /* Allow \octal to work the DWIM way (that is, stop scanning
425 * as soon as non-octal characters are seen, complain only if
426 * someone seems to want to use the digits eight and nine. Since we
427 * know it is not octal, then if isDIGIT, must be an 8 or 9). */
428 if (isDIGIT(*s)) {
429 if (!(*flags & PERL_SCAN_SILENT_ILLDIGIT))
430 Perl_ck_warner(aTHX_ packWARN(WARN_DIGIT),
431 "Illegal octal digit '%c' ignored", *s);
432 }
433 break;
434 }
435
436 if ( ( overflowed && value_nv > 4294967295.0)
437#if UVSIZE > 4
438 || (!overflowed && value > 0xffffffff
439 && ! (*flags & PERL_SCAN_SILENT_NON_PORTABLE))
440#endif
441 ) {
442 Perl_ck_warner(aTHX_ packWARN(WARN_PORTABLE),
443 "Octal number > 037777777777 non-portable");
444 }
445 *len_p = s - start;
446 if (!overflowed) {
447 *flags = 0;
448 return value;
449 }
450 *flags = PERL_SCAN_GREATER_THAN_UV_MAX;
451 if (result)
452 *result = value_nv;
453 return UV_MAX;
454}
455
456/*
457=for apidoc scan_bin
458
459For backwards compatibility. Use C<grok_bin> instead.
460
461=for apidoc scan_hex
462
463For backwards compatibility. Use C<grok_hex> instead.
464
465=for apidoc scan_oct
466
467For backwards compatibility. Use C<grok_oct> instead.
468
469=cut
470 */
471
472NV
473Perl_scan_bin(pTHX_ const char *start, STRLEN len, STRLEN *retlen)
474{
475 NV rnv;
476 I32 flags = *retlen ? PERL_SCAN_ALLOW_UNDERSCORES : 0;
477 const UV ruv = grok_bin (start, &len, &flags, &rnv);
478
479 PERL_ARGS_ASSERT_SCAN_BIN;
480
481 *retlen = len;
482 return (flags & PERL_SCAN_GREATER_THAN_UV_MAX) ? rnv : (NV)ruv;
483}
484
485NV
486Perl_scan_oct(pTHX_ const char *start, STRLEN len, STRLEN *retlen)
487{
488 NV rnv;
489 I32 flags = *retlen ? PERL_SCAN_ALLOW_UNDERSCORES : 0;
490 const UV ruv = grok_oct (start, &len, &flags, &rnv);
491
492 PERL_ARGS_ASSERT_SCAN_OCT;
493
494 *retlen = len;
495 return (flags & PERL_SCAN_GREATER_THAN_UV_MAX) ? rnv : (NV)ruv;
496}
497
498NV
499Perl_scan_hex(pTHX_ const char *start, STRLEN len, STRLEN *retlen)
500{
501 NV rnv;
502 I32 flags = *retlen ? PERL_SCAN_ALLOW_UNDERSCORES : 0;
503 const UV ruv = grok_hex (start, &len, &flags, &rnv);
504
505 PERL_ARGS_ASSERT_SCAN_HEX;
506
507 *retlen = len;
508 return (flags & PERL_SCAN_GREATER_THAN_UV_MAX) ? rnv : (NV)ruv;
509}
510
511/*
512=for apidoc grok_numeric_radix
513
514Scan and skip for a numeric decimal separator (radix).
515
516=cut
517 */
518bool
519Perl_grok_numeric_radix(pTHX_ const char **sp, const char *send)
520{
521#ifdef USE_LOCALE_NUMERIC
522 PERL_ARGS_ASSERT_GROK_NUMERIC_RADIX;
523
524 if (IN_LC(LC_NUMERIC)) {
525 DECLARATION_FOR_LC_NUMERIC_MANIPULATION;
526 STORE_LC_NUMERIC_SET_TO_NEEDED();
527 if (PL_numeric_radix_sv) {
528 STRLEN len;
529 const char * const radix = SvPV(PL_numeric_radix_sv, len);
530 if (*sp + len <= send && memEQ(*sp, radix, len)) {
531 *sp += len;
532 RESTORE_LC_NUMERIC();
533 return TRUE;
534 }
535 }
536 RESTORE_LC_NUMERIC();
537 }
538 /* always try "." if numeric radix didn't match because
539 * we may have data from different locales mixed */
540#endif
541
542 PERL_ARGS_ASSERT_GROK_NUMERIC_RADIX;
543
544 if (*sp < send && **sp == '.') {
545 ++*sp;
546 return TRUE;
547 }
548 return FALSE;
549}
550
551/*
552=for apidoc grok_infnan
553
554Helper for C<grok_number()>, accepts various ways of spelling "infinity"
555or "not a number", and returns one of the following flag combinations:
556
557 IS_NUMBER_INFINITE
558 IS_NUMBER_NAN
559 IS_NUMBER_INFINITE | IS_NUMBER_NEG
560 IS_NUMBER_NAN | IS_NUMBER_NEG
561 0
562
563possibly |-ed with C<IS_NUMBER_TRAILING>.
564
565If an infinity or a not-a-number is recognized, C<*sp> will point to
566one byte past the end of the recognized string. If the recognition fails,
567zero is returned, and C<*sp> will not move.
568
569=cut
570*/
571
572int
573Perl_grok_infnan(pTHX_ const char** sp, const char* send)
574{
575 const char* s = *sp;
576 int flags = 0;
577#if defined(NV_INF) || defined(NV_NAN)
578 bool odh = FALSE; /* one-dot-hash: 1.#INF */
579
580 PERL_ARGS_ASSERT_GROK_INFNAN;
581
582 if (*s == '+') {
583 s++; if (s == send) return 0;
584 }
585 else if (*s == '-') {
586 flags |= IS_NUMBER_NEG; /* Yes, -NaN happens. Incorrect but happens. */
587 s++; if (s == send) return 0;
588 }
589
590 if (*s == '1') {
591 /* Visual C: 1.#SNAN, -1.#QNAN, 1#INF, 1.#IND (maybe also 1.#NAN)
592 * Let's keep the dot optional. */
593 s++; if (s == send) return 0;
594 if (*s == '.') {
595 s++; if (s == send) return 0;
596 }
597 if (*s == '#') {
598 s++; if (s == send) return 0;
599 } else
600 return 0;
601 odh = TRUE;
602 }
603
604 if (isALPHA_FOLD_EQ(*s, 'I')) {
605 /* INF or IND (1.#IND is "indeterminate", a certain type of NAN) */
606
607 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
608 s++; if (s == send) return 0;
609 if (isALPHA_FOLD_EQ(*s, 'F')) {
610 s++;
611 if (s < send && (isALPHA_FOLD_EQ(*s, 'I'))) {
612 int fail =
613 flags | IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT | IS_NUMBER_TRAILING;
614 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return fail;
615 s++; if (s == send || isALPHA_FOLD_NE(*s, 'I')) return fail;
616 s++; if (s == send || isALPHA_FOLD_NE(*s, 'T')) return fail;
617 s++; if (s == send || isALPHA_FOLD_NE(*s, 'Y')) return fail;
618 s++;
619 } else if (odh) {
620 while (*s == '0') { /* 1.#INF00 */
621 s++;
622 }
623 }
624 while (s < send && isSPACE(*s))
625 s++;
626 if (s < send && *s) {
627 flags |= IS_NUMBER_TRAILING;
628 }
629 flags |= IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT;
630 }
631 else if (isALPHA_FOLD_EQ(*s, 'D') && odh) { /* 1.#IND */
632 s++;
633 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
634 while (*s == '0') { /* 1.#IND00 */
635 s++;
636 }
637 if (*s) {
638 flags |= IS_NUMBER_TRAILING;
639 }
640 } else
641 return 0;
642 }
643 else {
644 /* Maybe NAN of some sort */
645
646 if (isALPHA_FOLD_EQ(*s, 'S') || isALPHA_FOLD_EQ(*s, 'Q')) {
647 /* snan, qNaN */
648 /* XXX do something with the snan/qnan difference */
649 s++; if (s == send) return 0;
650 }
651
652 if (isALPHA_FOLD_EQ(*s, 'N')) {
653 s++; if (s == send || isALPHA_FOLD_NE(*s, 'A')) return 0;
654 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
655 s++;
656
657 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
658
659 /* NaN can be followed by various stuff (NaNQ, NaNS), but
660 * there are also multiple different NaN values, and some
661 * implementations output the "payload" values,
662 * e.g. NaN123, NAN(abc), while some legacy implementations
663 * have weird stuff like NaN%. */
664 if (isALPHA_FOLD_EQ(*s, 'q') ||
665 isALPHA_FOLD_EQ(*s, 's')) {
666 /* "nanq" or "nans" are ok, though generating
667 * these portably is tricky. */
668 s++;
669 }
670 if (*s == '(') {
671 /* C99 style "nan(123)" or Perlish equivalent "nan($uv)". */
672 const char *t;
673 s++;
674 if (s == send) {
675 return flags | IS_NUMBER_TRAILING;
676 }
677 t = s + 1;
678 while (t < send && *t && *t != ')') {
679 t++;
680 }
681 if (t == send) {
682 return flags | IS_NUMBER_TRAILING;
683 }
684 if (*t == ')') {
685 int nantype;
686 UV nanval;
687 if (s[0] == '0' && s + 2 < t &&
688 isALPHA_FOLD_EQ(s[1], 'x') &&
689 isXDIGIT(s[2])) {
690 STRLEN len = t - s;
691 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES;
692 nanval = grok_hex(s, &len, &flags, NULL);
693 if ((flags & PERL_SCAN_GREATER_THAN_UV_MAX)) {
694 nantype = 0;
695 } else {
696 nantype = IS_NUMBER_IN_UV;
697 }
698 s += len;
699 } else if (s[0] == '0' && s + 2 < t &&
700 isALPHA_FOLD_EQ(s[1], 'b') &&
701 (s[2] == '0' || s[2] == '1')) {
702 STRLEN len = t - s;
703 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES;
704 nanval = grok_bin(s, &len, &flags, NULL);
705 if ((flags & PERL_SCAN_GREATER_THAN_UV_MAX)) {
706 nantype = 0;
707 } else {
708 nantype = IS_NUMBER_IN_UV;
709 }
710 s += len;
711 } else {
712 const char *u;
713 nantype =
714 grok_number_flags(s, t - s, &nanval,
715 PERL_SCAN_TRAILING |
716 PERL_SCAN_ALLOW_UNDERSCORES);
717 /* Unfortunately grok_number_flags() doesn't
718 * tell how far we got and the ')' will always
719 * be "trailing", so we need to double-check
720 * whether we had something dubious. */
721 for (u = s; u < t; u++) {
722 if (!isDIGIT(*u)) {
723 flags |= IS_NUMBER_TRAILING;
724 break;
725 }
726 }
727 s = u;
728 }
729
730 /* XXX Doesn't do octal: nan("0123").
731 * Probably not a big loss. */
732
733 if ((nantype & IS_NUMBER_NOT_INT) ||
734 !(nantype && IS_NUMBER_IN_UV)) {
735 /* XXX the nanval is currently unused, that is,
736 * not inserted as the NaN payload of the NV.
737 * But the above code already parses the C99
738 * nan(...) format. See below, and see also
739 * the nan() in POSIX.xs.
740 *
741 * Certain configuration combinations where
742 * NVSIZE is greater than UVSIZE mean that
743 * a single UV cannot contain all the possible
744 * NaN payload bits. There would need to be
745 * some more generic syntax than "nan($uv)".
746 *
747 * Issues to keep in mind:
748 *
749 * (1) In most common cases there would
750 * not be an integral number of bytes that
751 * could be set, only a certain number of bits.
752 * For example for the common case of
753 * NVSIZE == UVSIZE == 8 there is room for 52
754 * bits in the payload, but the most significant
755 * bit is commonly reserved for the
756 * signaling/quiet bit, leaving 51 bits.
757 * Furthermore, the C99 nan() is supposed
758 * to generate quiet NaNs, so it is doubtful
759 * whether it should be able to generate
760 * signaling NaNs. For the x86 80-bit doubles
761 * (if building a long double Perl) there would
762 * be 62 bits (s/q bit being the 63rd).
763 *
764 * (2) Endianness of the payload bits. If the
765 * payload is specified as an UV, the low-order
766 * bits of the UV are naturally little-endianed
767 * (rightmost) bits of the payload. The endianness
768 * of UVs and NVs can be different. */
769 return 0;
770 }
771 if (s < t) {
772 flags |= IS_NUMBER_TRAILING;
773 }
774 } else {
775 /* Looked like nan(...), but no close paren. */
776 flags |= IS_NUMBER_TRAILING;
777 }
778 } else {
779 while (s < send && isSPACE(*s))
780 s++;
781 if (s < send && *s) {
782 /* Note that we here implicitly accept (parse as
783 * "nan", but with warnings) also any other weird
784 * trailing stuff for "nan". In the above we just
785 * check that if we got the C99-style "nan(...)",
786 * the "..." looks sane.
787 * If in future we accept more ways of specifying
788 * the nan payload, the accepting would happen around
789 * here. */
790 flags |= IS_NUMBER_TRAILING;
791 }
792 }
793 s = send;
794 }
795 else
796 return 0;
797 }
798
799 while (s < send && isSPACE(*s))
800 s++;
801
802#else
803 PERL_UNUSED_ARG(send);
804#endif /* #if defined(NV_INF) || defined(NV_NAN) */
805 *sp = s;
806 return flags;
807}
808
809/*
810=for apidoc grok_number_flags
811
812Recognise (or not) a number. The type of the number is returned
813(0 if unrecognised), otherwise it is a bit-ORed combination of
814C<IS_NUMBER_IN_UV>, C<IS_NUMBER_GREATER_THAN_UV_MAX>, C<IS_NUMBER_NOT_INT>,
815C<IS_NUMBER_NEG>, C<IS_NUMBER_INFINITY>, C<IS_NUMBER_NAN> (defined in perl.h).
816
817If the value of the number can fit in a UV, it is returned in C<*valuep>.
818C<IS_NUMBER_IN_UV> will be set to indicate that C<*valuep> is valid, C<IS_NUMBER_IN_UV>
819will never be set unless C<*valuep> is valid, but C<*valuep> may have been assigned
820to during processing even though C<IS_NUMBER_IN_UV> is not set on return.
821If C<valuep> is C<NULL>, C<IS_NUMBER_IN_UV> will be set for the same cases as when
822C<valuep> is non-C<NULL>, but no actual assignment (or SEGV) will occur.
823
824C<IS_NUMBER_NOT_INT> will be set with C<IS_NUMBER_IN_UV> if trailing decimals were
825seen (in which case C<*valuep> gives the true value truncated to an integer), and
826C<IS_NUMBER_NEG> if the number is negative (in which case C<*valuep> holds the
827absolute value). C<IS_NUMBER_IN_UV> is not set if e notation was used or the
828number is larger than a UV.
829
830C<flags> allows only C<PERL_SCAN_TRAILING>, which allows for trailing
831non-numeric text on an otherwise successful I<grok>, setting
832C<IS_NUMBER_TRAILING> on the result.
833
834=for apidoc grok_number
835
836Identical to C<grok_number_flags()> with C<flags> set to zero.
837
838=cut
839 */
840int
841Perl_grok_number(pTHX_ const char *pv, STRLEN len, UV *valuep)
842{
843 PERL_ARGS_ASSERT_GROK_NUMBER;
844
845 return grok_number_flags(pv, len, valuep, 0);
846}
847
848static const UV uv_max_div_10 = UV_MAX / 10;
849static const U8 uv_max_mod_10 = UV_MAX % 10;
850
851int
852Perl_grok_number_flags(pTHX_ const char *pv, STRLEN len, UV *valuep, U32 flags)
853{
854 const char *s = pv;
855 const char * const send = pv + len;
856 const char *d;
857 int numtype = 0;
858
859 PERL_ARGS_ASSERT_GROK_NUMBER_FLAGS;
860
861 while (s < send && isSPACE(*s))
862 s++;
863 if (s == send) {
864 return 0;
865 } else if (*s == '-') {
866 s++;
867 numtype = IS_NUMBER_NEG;
868 }
869 else if (*s == '+')
870 s++;
871
872 if (s == send)
873 return 0;
874
875 /* The first digit (after optional sign): note that might
876 * also point to "infinity" or "nan", or "1.#INF". */
877 d = s;
878
879 /* next must be digit or the radix separator or beginning of infinity/nan */
880 if (isDIGIT(*s)) {
881 /* UVs are at least 32 bits, so the first 9 decimal digits cannot
882 overflow. */
883 UV value = *s - '0';
884 /* This construction seems to be more optimiser friendly.
885 (without it gcc does the isDIGIT test and the *s - '0' separately)
886 With it gcc on arm is managing 6 instructions (6 cycles) per digit.
887 In theory the optimiser could deduce how far to unroll the loop
888 before checking for overflow. */
889 if (++s < send) {
890 int digit = *s - '0';
891 if (digit >= 0 && digit <= 9) {
892 value = value * 10 + digit;
893 if (++s < send) {
894 digit = *s - '0';
895 if (digit >= 0 && digit <= 9) {
896 value = value * 10 + digit;
897 if (++s < send) {
898 digit = *s - '0';
899 if (digit >= 0 && digit <= 9) {
900 value = value * 10 + digit;
901 if (++s < send) {
902 digit = *s - '0';
903 if (digit >= 0 && digit <= 9) {
904 value = value * 10 + digit;
905 if (++s < send) {
906 digit = *s - '0';
907 if (digit >= 0 && digit <= 9) {
908 value = value * 10 + digit;
909 if (++s < send) {
910 digit = *s - '0';
911 if (digit >= 0 && digit <= 9) {
912 value = value * 10 + digit;
913 if (++s < send) {
914 digit = *s - '0';
915 if (digit >= 0 && digit <= 9) {
916 value = value * 10 + digit;
917 if (++s < send) {
918 digit = *s - '0';
919 if (digit >= 0 && digit <= 9) {
920 value = value * 10 + digit;
921 if (++s < send) {
922 /* Now got 9 digits, so need to check
923 each time for overflow. */
924 digit = *s - '0';
925 while (digit >= 0 && digit <= 9
926 && (value < uv_max_div_10
927 || (value == uv_max_div_10
928 && digit <= uv_max_mod_10))) {
929 value = value * 10 + digit;
930 if (++s < send)
931 digit = *s - '0';
932 else
933 break;
934 }
935 if (digit >= 0 && digit <= 9
936 && (s < send)) {
937 /* value overflowed.
938 skip the remaining digits, don't
939 worry about setting *valuep. */
940 do {
941 s++;
942 } while (s < send && isDIGIT(*s));
943 numtype |=
944 IS_NUMBER_GREATER_THAN_UV_MAX;
945 goto skip_value;
946 }
947 }
948 }
949 }
950 }
951 }
952 }
953 }
954 }
955 }
956 }
957 }
958 }
959 }
960 }
961 }
962 }
963 }
964 numtype |= IS_NUMBER_IN_UV;
965 if (valuep)
966 *valuep = value;
967
968 skip_value:
969 if (GROK_NUMERIC_RADIX(&s, send)) {
970 numtype |= IS_NUMBER_NOT_INT;
971 while (s < send && isDIGIT(*s)) /* optional digits after the radix */
972 s++;
973 }
974 }
975 else if (GROK_NUMERIC_RADIX(&s, send)) {
976 numtype |= IS_NUMBER_NOT_INT | IS_NUMBER_IN_UV; /* valuep assigned below */
977 /* no digits before the radix means we need digits after it */
978 if (s < send && isDIGIT(*s)) {
979 do {
980 s++;
981 } while (s < send && isDIGIT(*s));
982 if (valuep) {
983 /* integer approximation is valid - it's 0. */
984 *valuep = 0;
985 }
986 }
987 else
988 return 0;
989 }
990
991 if (s > d && s < send) {
992 /* we can have an optional exponent part */
993 if (isALPHA_FOLD_EQ(*s, 'e')) {
994 s++;
995 if (s < send && (*s == '-' || *s == '+'))
996 s++;
997 if (s < send && isDIGIT(*s)) {
998 do {
999 s++;
1000 } while (s < send && isDIGIT(*s));
1001 }
1002 else if (flags & PERL_SCAN_TRAILING)
1003 return numtype | IS_NUMBER_TRAILING;
1004 else
1005 return 0;
1006
1007 /* The only flag we keep is sign. Blow away any "it's UV" */
1008 numtype &= IS_NUMBER_NEG;
1009 numtype |= IS_NUMBER_NOT_INT;
1010 }
1011 }
1012 while (s < send && isSPACE(*s))
1013 s++;
1014 if (s >= send)
1015 return numtype;
1016 if (len == 10 && _memEQs(pv, "0 but true")) {
1017 if (valuep)
1018 *valuep = 0;
1019 return IS_NUMBER_IN_UV;
1020 }
1021 /* We could be e.g. at "Inf" or "NaN", or at the "#" of "1.#INF". */
1022 if ((s + 2 < send) && strchr("inqs#", toFOLD(*s))) {
1023 /* Really detect inf/nan. Start at d, not s, since the above
1024 * code might have already consumed the "1." or "1". */
1025 const int infnan = Perl_grok_infnan(aTHX_ &d, send);
1026 if ((infnan & IS_NUMBER_INFINITY)) {
1027 return (numtype | infnan); /* Keep sign for infinity. */
1028 }
1029 else if ((infnan & IS_NUMBER_NAN)) {
1030 return (numtype | infnan) & ~IS_NUMBER_NEG; /* Clear sign for nan. */
1031 }
1032 }
1033 else if (flags & PERL_SCAN_TRAILING) {
1034 return numtype | IS_NUMBER_TRAILING;
1035 }
1036
1037 return 0;
1038}
1039
1040/*
1041grok_atoUV
1042
1043grok_atoUV parses a C-style zero-byte terminated string, looking for
1044a decimal unsigned integer.
1045
1046Returns the unsigned integer, if a valid value can be parsed
1047from the beginning of the string.
1048
1049Accepts only the decimal digits '0'..'9'.
1050
1051As opposed to atoi or strtol, grok_atoUV does NOT allow optional
1052leading whitespace, or negative inputs. If such features are
1053required, the calling code needs to explicitly implement those.
1054
1055Returns true if a valid value could be parsed. In that case, valptr
1056is set to the parsed value, and endptr (if provided) is set to point
1057to the character after the last digit.
1058
1059Returns false otherwise. This can happen if a) there is a leading zero
1060followed by another digit; b) the digits would overflow a UV; or c)
1061there are trailing non-digits AND endptr is not provided.
1062
1063Background: atoi has severe problems with illegal inputs, it cannot be
1064used for incremental parsing, and therefore should be avoided
1065atoi and strtol are also affected by locale settings, which can also be
1066seen as a bug (global state controlled by user environment).
1067
1068*/
1069
1070bool
1071Perl_grok_atoUV(const char *pv, UV *valptr, const char** endptr)
1072{
1073 const char* s = pv;
1074 const char** eptr;
1075 const char* end2; /* Used in case endptr is NULL. */
1076 UV val = 0; /* The parsed value. */
1077
1078 PERL_ARGS_ASSERT_GROK_ATOUV;
1079
1080 eptr = endptr ? endptr : &end2;
1081 if (isDIGIT(*s)) {
1082 /* Single-digit inputs are quite common. */
1083 val = *s++ - '0';
1084 if (isDIGIT(*s)) {
1085 /* Fail on extra leading zeros. */
1086 if (val == 0)
1087 return FALSE;
1088 while (isDIGIT(*s)) {
1089 /* This could be unrolled like in grok_number(), but
1090 * the expected uses of this are not speed-needy, and
1091 * unlikely to need full 64-bitness. */
1092 const U8 digit = *s++ - '0';
1093 if (val < uv_max_div_10 ||
1094 (val == uv_max_div_10 && digit <= uv_max_mod_10)) {
1095 val = val * 10 + digit;
1096 } else {
1097 return FALSE;
1098 }
1099 }
1100 }
1101 }
1102 if (s == pv)
1103 return FALSE;
1104 if (endptr == NULL && *s)
1105 return FALSE; /* If endptr is NULL, no trailing non-digits allowed. */
1106 *eptr = s;
1107 *valptr = val;
1108 return TRUE;
1109}
1110
1111#ifndef USE_QUADMATH
1112STATIC NV
1113S_mulexp10(NV value, I32 exponent)
1114{
1115 NV result = 1.0;
1116 NV power = 10.0;
1117 bool negative = 0;
1118 I32 bit;
1119
1120 if (exponent == 0)
1121 return value;
1122 if (value == 0)
1123 return (NV)0;
1124
1125 /* On OpenVMS VAX we by default use the D_FLOAT double format,
1126 * and that format does not have *easy* capabilities [1] for
1127 * overflowing doubles 'silently' as IEEE fp does. We also need
1128 * to support G_FLOAT on both VAX and Alpha, and though the exponent
1129 * range is much larger than D_FLOAT it still doesn't do silent
1130 * overflow. Therefore we need to detect early whether we would
1131 * overflow (this is the behaviour of the native string-to-float
1132 * conversion routines, and therefore of native applications, too).
1133 *
1134 * [1] Trying to establish a condition handler to trap floating point
1135 * exceptions is not a good idea. */
1136
1137 /* In UNICOS and in certain Cray models (such as T90) there is no
1138 * IEEE fp, and no way at all from C to catch fp overflows gracefully.
1139 * There is something you can do if you are willing to use some
1140 * inline assembler: the instruction is called DFI-- but that will
1141 * disable *all* floating point interrupts, a little bit too large
1142 * a hammer. Therefore we need to catch potential overflows before
1143 * it's too late. */
1144
1145#if ((defined(VMS) && !defined(_IEEE_FP)) || defined(_UNICOS) || defined(DOUBLE_IS_VAX_FLOAT)) && defined(NV_MAX_10_EXP)
1146 STMT_START {
1147 const NV exp_v = log10(value);
1148 if (exponent >= NV_MAX_10_EXP || exponent + exp_v >= NV_MAX_10_EXP)
1149 return NV_MAX;
1150 if (exponent < 0) {
1151 if (-(exponent + exp_v) >= NV_MAX_10_EXP)
1152 return 0.0;
1153 while (-exponent >= NV_MAX_10_EXP) {
1154 /* combination does not overflow, but 10^(-exponent) does */
1155 value /= 10;
1156 ++exponent;
1157 }
1158 }
1159 } STMT_END;
1160#endif
1161
1162 if (exponent < 0) {
1163 negative = 1;
1164 exponent = -exponent;
1165#ifdef NV_MAX_10_EXP
1166 /* for something like 1234 x 10^-309, the action of calculating
1167 * the intermediate value 10^309 then returning 1234 / (10^309)
1168 * will fail, since 10^309 becomes infinity. In this case try to
1169 * refactor it as 123 / (10^308) etc.
1170 */
1171 while (value && exponent > NV_MAX_10_EXP) {
1172 exponent--;
1173 value /= 10;
1174 }
1175 if (value == 0.0)
1176 return value;
1177#endif
1178 }
1179#if defined(__osf__)
1180 /* Even with cc -ieee + ieee_set_fp_control(IEEE_TRAP_ENABLE_INV)
1181 * Tru64 fp behavior on inf/nan is somewhat broken. Another way
1182 * to do this would be ieee_set_fp_control(IEEE_TRAP_ENABLE_OVF)
1183 * but that breaks another set of infnan.t tests. */
1184# define FP_OVERFLOWS_TO_ZERO
1185#endif
1186 for (bit = 1; exponent; bit <<= 1) {
1187 if (exponent & bit) {
1188 exponent ^= bit;
1189 result *= power;
1190#ifdef FP_OVERFLOWS_TO_ZERO
1191 if (result == 0)
1192# ifdef NV_INF
1193 return value < 0 ? -NV_INF : NV_INF;
1194# else
1195 return value < 0 ? -FLT_MAX : FLT_MAX;
1196# endif
1197#endif
1198 /* Floating point exceptions are supposed to be turned off,
1199 * but if we're obviously done, don't risk another iteration.
1200 */
1201 if (exponent == 0) break;
1202 }
1203 power *= power;
1204 }
1205 return negative ? value / result : value * result;
1206}
1207#endif /* #ifndef USE_QUADMATH */
1208
1209NV
1210Perl_my_atof(pTHX_ const char* s)
1211{
1212 NV x = 0.0;
1213#ifdef USE_QUADMATH
1214 Perl_my_atof2(aTHX_ s, &x);
1215 return x;
1216#else
1217# ifdef USE_LOCALE_NUMERIC
1218 PERL_ARGS_ASSERT_MY_ATOF;
1219
1220 {
1221 DECLARATION_FOR_LC_NUMERIC_MANIPULATION;
1222 STORE_LC_NUMERIC_SET_TO_NEEDED();
1223 if (PL_numeric_radix_sv && IN_LC(LC_NUMERIC)) {
1224 /* Look through the string for the first thing that looks like a
1225 * decimal point: either the value in the current locale or the
1226 * standard fallback of '.'. The one which appears earliest in the
1227 * input string is the one that we should have atof look for. Note
1228 * that we have to determine this beforehand because on some
1229 * systems, Perl_atof2 is just a wrapper around the system's atof.
1230 * */
1231 const char * const standard = strchr(s, '.');
1232 const char * const local = strstr(s, SvPV_nolen(PL_numeric_radix_sv));
1233 const bool use_standard_radix = standard && (!local || standard < local);
1234
1235 if (use_standard_radix)
1236 SET_NUMERIC_STANDARD();
1237
1238 Perl_atof2(s, x);
1239
1240 if (use_standard_radix)
1241 SET_NUMERIC_UNDERLYING();
1242 }
1243 else
1244 Perl_atof2(s, x);
1245 RESTORE_LC_NUMERIC();
1246 }
1247# else
1248 Perl_atof2(s, x);
1249# endif
1250#endif
1251 return x;
1252}
1253
1254#if defined(NV_INF) || defined(NV_NAN)
1255
1256#ifdef USING_MSVC6
1257# pragma warning(push)
1258# pragma warning(disable:4756;disable:4056)
1259#endif
1260static char*
1261S_my_atof_infnan(pTHX_ const char* s, bool negative, const char* send, NV* value)
1262{
1263 const char *p0 = negative ? s - 1 : s;
1264 const char *p = p0;
1265 const int infnan = grok_infnan(&p, send);
1266 if (infnan && p != p0) {
1267 /* If we can generate inf/nan directly, let's do so. */
1268#ifdef NV_INF
1269 if ((infnan & IS_NUMBER_INFINITY)) {
1270 *value = (infnan & IS_NUMBER_NEG) ? -NV_INF: NV_INF;
1271 return (char*)p;
1272 }
1273#endif
1274#ifdef NV_NAN
1275 if ((infnan & IS_NUMBER_NAN)) {
1276 *value = NV_NAN;
1277 return (char*)p;
1278 }
1279#endif
1280#ifdef Perl_strtod
1281 /* If still here, we didn't have either NV_INF or NV_NAN,
1282 * and can try falling back to native strtod/strtold.
1283 *
1284 * The native interface might not recognize all the possible
1285 * inf/nan strings Perl recognizes. What we can try
1286 * is to try faking the input. We will try inf/-inf/nan
1287 * as the most promising/portable input. */
1288 {
1289 const char* fake = NULL;
1290 char* endp;
1291 NV nv;
1292#ifdef NV_INF
1293 if ((infnan & IS_NUMBER_INFINITY)) {
1294 fake = ((infnan & IS_NUMBER_NEG)) ? "-inf" : "inf";
1295 }
1296#endif
1297#ifdef NV_NAN
1298 if ((infnan & IS_NUMBER_NAN)) {
1299 fake = "nan";
1300 }
1301#endif
1302 assert(fake);
1303 nv = Perl_strtod(fake, &endp);
1304 if (fake != endp) {
1305#ifdef NV_INF
1306 if ((infnan & IS_NUMBER_INFINITY)) {
1307# ifdef Perl_isinf
1308 if (Perl_isinf(nv))
1309 *value = nv;
1310# else
1311 /* last resort, may generate SIGFPE */
1312 *value = Perl_exp((NV)1e9);
1313 if ((infnan & IS_NUMBER_NEG))
1314 *value = -*value;
1315# endif
1316 return (char*)p; /* p, not endp */
1317 }
1318#endif
1319#ifdef NV_NAN
1320 if ((infnan & IS_NUMBER_NAN)) {
1321# ifdef Perl_isnan
1322 if (Perl_isnan(nv))
1323 *value = nv;
1324# else
1325 /* last resort, may generate SIGFPE */
1326 *value = Perl_log((NV)-1.0);
1327# endif
1328 return (char*)p; /* p, not endp */
1329#endif
1330 }
1331 }
1332 }
1333#endif /* #ifdef Perl_strtod */
1334 }
1335 return NULL;
1336}
1337#ifdef USING_MSVC6
1338# pragma warning(pop)
1339#endif
1340
1341#endif /* if defined(NV_INF) || defined(NV_NAN) */
1342
1343char*
1344Perl_my_atof2(pTHX_ const char* orig, NV* value)
1345{
1346 const char* s = orig;
1347 NV result[3] = {0.0, 0.0, 0.0};
1348#if defined(USE_PERL_ATOF) || defined(USE_QUADMATH)
1349 const char* send = s + strlen(orig); /* one past the last */
1350 bool negative = 0;
1351#endif
1352#if defined(USE_PERL_ATOF) && !defined(USE_QUADMATH)
1353 UV accumulator[2] = {0,0}; /* before/after dp */
1354 bool seen_digit = 0;
1355 I32 exp_adjust[2] = {0,0};
1356 I32 exp_acc[2] = {-1, -1};
1357 /* the current exponent adjust for the accumulators */
1358 I32 exponent = 0;
1359 I32 seen_dp = 0;
1360 I32 digit = 0;
1361 I32 old_digit = 0;
1362 I32 sig_digits = 0; /* noof significant digits seen so far */
1363#endif
1364
1365#if defined(USE_PERL_ATOF) || defined(USE_QUADMATH)
1366 PERL_ARGS_ASSERT_MY_ATOF2;
1367
1368 /* leading whitespace */
1369 while (isSPACE(*s))
1370 ++s;
1371
1372 /* sign */
1373 switch (*s) {
1374 case '-':
1375 negative = 1;
1376 /* FALLTHROUGH */
1377 case '+':
1378 ++s;
1379 }
1380#endif
1381
1382#ifdef USE_QUADMATH
1383 {
1384 char* endp;
1385 if ((endp = S_my_atof_infnan(aTHX_ s, negative, send, value)))
1386 return endp;
1387 result[2] = strtoflt128(s, &endp);
1388 if (s != endp) {
1389 *value = negative ? -result[2] : result[2];
1390 return endp;
1391 }
1392 return NULL;
1393 }
1394#elif defined(USE_PERL_ATOF)
1395
1396/* There is no point in processing more significant digits
1397 * than the NV can hold. Note that NV_DIG is a lower-bound value,
1398 * while we need an upper-bound value. We add 2 to account for this;
1399 * since it will have been conservative on both the first and last digit.
1400 * For example a 32-bit mantissa with an exponent of 4 would have
1401 * exact values in the set
1402 * 4
1403 * 8
1404 * ..
1405 * 17179869172
1406 * 17179869176
1407 * 17179869180
1408 *
1409 * where for the purposes of calculating NV_DIG we would have to discount
1410 * both the first and last digit, since neither can hold all values from
1411 * 0..9; but for calculating the value we must examine those two digits.
1412 */
1413#ifdef MAX_SIG_DIG_PLUS
1414 /* It is not necessarily the case that adding 2 to NV_DIG gets all the
1415 possible digits in a NV, especially if NVs are not IEEE compliant
1416 (e.g., long doubles on IRIX) - Allen <allens@cpan.org> */
1417# define MAX_SIG_DIGITS (NV_DIG+MAX_SIG_DIG_PLUS)
1418#else
1419# define MAX_SIG_DIGITS (NV_DIG+2)
1420#endif
1421
1422/* the max number we can accumulate in a UV, and still safely do 10*N+9 */
1423#define MAX_ACCUMULATE ( (UV) ((UV_MAX - 9)/10))
1424
1425#if defined(NV_INF) || defined(NV_NAN)
1426 {
1427 char* endp;
1428 if ((endp = S_my_atof_infnan(aTHX_ s, negative, send, value)))
1429 return endp;
1430 }
1431#endif
1432
1433 /* we accumulate digits into an integer; when this becomes too
1434 * large, we add the total to NV and start again */
1435
1436 while (1) {
1437 if (isDIGIT(*s)) {
1438 seen_digit = 1;
1439 old_digit = digit;
1440 digit = *s++ - '0';
1441 if (seen_dp)
1442 exp_adjust[1]++;
1443
1444 /* don't start counting until we see the first significant
1445 * digit, eg the 5 in 0.00005... */
1446 if (!sig_digits && digit == 0)
1447 continue;
1448
1449 if (++sig_digits > MAX_SIG_DIGITS) {
1450 /* limits of precision reached */
1451 if (digit > 5) {
1452 ++accumulator[seen_dp];
1453 } else if (digit == 5) {
1454 if (old_digit % 2) { /* round to even - Allen */
1455 ++accumulator[seen_dp];
1456 }
1457 }
1458 if (seen_dp) {
1459 exp_adjust[1]--;
1460 } else {
1461 exp_adjust[0]++;
1462 }
1463 /* skip remaining digits */
1464 while (isDIGIT(*s)) {
1465 ++s;
1466 if (! seen_dp) {
1467 exp_adjust[0]++;
1468 }
1469 }
1470 /* warn of loss of precision? */
1471 }
1472 else {
1473 if (accumulator[seen_dp] > MAX_ACCUMULATE) {
1474 /* add accumulator to result and start again */
1475 result[seen_dp] = S_mulexp10(result[seen_dp],
1476 exp_acc[seen_dp])
1477 + (NV)accumulator[seen_dp];
1478 accumulator[seen_dp] = 0;
1479 exp_acc[seen_dp] = 0;
1480 }
1481 accumulator[seen_dp] = accumulator[seen_dp] * 10 + digit;
1482 ++exp_acc[seen_dp];
1483 }
1484 }
1485 else if (!seen_dp && GROK_NUMERIC_RADIX(&s, send)) {
1486 seen_dp = 1;
1487 if (sig_digits > MAX_SIG_DIGITS) {
1488 do {
1489 ++s;
1490 } while (isDIGIT(*s));
1491 break;
1492 }
1493 }
1494 else {
1495 break;
1496 }
1497 }
1498
1499 result[0] = S_mulexp10(result[0], exp_acc[0]) + (NV)accumulator[0];
1500 if (seen_dp) {
1501 result[1] = S_mulexp10(result[1], exp_acc[1]) + (NV)accumulator[1];
1502 }
1503
1504 if (seen_digit && (isALPHA_FOLD_EQ(*s, 'e'))) {
1505 bool expnegative = 0;
1506
1507 ++s;
1508 switch (*s) {
1509 case '-':
1510 expnegative = 1;
1511 /* FALLTHROUGH */
1512 case '+':
1513 ++s;
1514 }
1515 while (isDIGIT(*s))
1516 exponent = exponent * 10 + (*s++ - '0');
1517 if (expnegative)
1518 exponent = -exponent;
1519 }
1520
1521
1522
1523 /* now apply the exponent */
1524
1525 if (seen_dp) {
1526 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0])
1527 + S_mulexp10(result[1],exponent-exp_adjust[1]);
1528 } else {
1529 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0]);
1530 }
1531
1532 /* now apply the sign */
1533 if (negative)
1534 result[2] = -result[2];
1535#endif /* USE_PERL_ATOF */
1536 *value = result[2];
1537 return (char *)s;
1538}
1539
1540/*
1541=for apidoc isinfnan
1542
1543C<Perl_isinfnan()> is utility function that returns true if the NV
1544argument is either an infinity or a C<NaN>, false otherwise. To test
1545in more detail, use C<Perl_isinf()> and C<Perl_isnan()>.
1546
1547This is also the logical inverse of Perl_isfinite().
1548
1549=cut
1550*/
1551bool
1552Perl_isinfnan(NV nv)
1553{
1554 PERL_UNUSED_ARG(nv);
1555#ifdef Perl_isinf
1556 if (Perl_isinf(nv))
1557 return TRUE;
1558#endif
1559#ifdef Perl_isnan
1560 if (Perl_isnan(nv))
1561 return TRUE;
1562#endif
1563 return FALSE;
1564}
1565
1566/*
1567=for apidoc
1568
1569Checks whether the argument would be either an infinity or C<NaN> when used
1570as a number, but is careful not to trigger non-numeric or uninitialized
1571warnings. it assumes the caller has done C<SvGETMAGIC(sv)> already.
1572
1573=cut
1574*/
1575
1576bool
1577Perl_isinfnansv(pTHX_ SV *sv)
1578{
1579 PERL_ARGS_ASSERT_ISINFNANSV;
1580 if (!SvOK(sv))
1581 return FALSE;
1582 if (SvNOKp(sv))
1583 return Perl_isinfnan(SvNVX(sv));
1584 if (SvIOKp(sv))
1585 return FALSE;
1586 {
1587 STRLEN len;
1588 const char *s = SvPV_nomg_const(sv, len);
1589 return cBOOL(grok_infnan(&s, s+len));
1590 }
1591}
1592
1593#ifndef HAS_MODFL
1594/* C99 has truncl, pre-C99 Solaris had aintl. We can use either with
1595 * copysignl to emulate modfl, which is in some platforms missing or
1596 * broken. */
1597# if defined(HAS_TRUNCL) && defined(HAS_COPYSIGNL)
1598long double
1599Perl_my_modfl(long double x, long double *ip)
1600{
1601 *ip = truncl(x);
1602 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
1603}
1604# elif defined(HAS_AINTL) && defined(HAS_COPYSIGNL)
1605long double
1606Perl_my_modfl(long double x, long double *ip)
1607{
1608 *ip = aintl(x);
1609 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
1610}
1611# endif
1612#endif
1613
1614/* Similarly, with ilogbl and scalbnl we can emulate frexpl. */
1615#if ! defined(HAS_FREXPL) && defined(HAS_ILOGBL) && defined(HAS_SCALBNL)
1616long double
1617Perl_my_frexpl(long double x, int *e) {
1618 *e = x == 0.0L ? 0 : ilogbl(x) + 1;
1619 return (scalbnl(x, -*e));
1620}
1621#endif
1622
1623/*
1624=for apidoc Perl_signbit
1625
1626Return a non-zero integer if the sign bit on an NV is set, and 0 if
1627it is not.
1628
1629If F<Configure> detects this system has a C<signbit()> that will work with
1630our NVs, then we just use it via the C<#define> in F<perl.h>. Otherwise,
1631fall back on this implementation. The main use of this function
1632is catching C<-0.0>.
1633
1634C<Configure> notes: This function is called C<'Perl_signbit'> instead of a
1635plain C<'signbit'> because it is easy to imagine a system having a C<signbit()>
1636function or macro that doesn't happen to work with our particular choice
1637of NVs. We shouldn't just re-C<#define> C<signbit> as C<Perl_signbit> and expect
1638the standard system headers to be happy. Also, this is a no-context
1639function (no C<pTHX_>) because C<Perl_signbit()> is usually re-C<#defined> in
1640F<perl.h> as a simple macro call to the system's C<signbit()>.
1641Users should just always call C<Perl_signbit()>.
1642
1643=cut
1644*/
1645#if !defined(HAS_SIGNBIT)
1646int
1647Perl_signbit(NV x) {
1648# ifdef Perl_fp_class_nzero
1649 return Perl_fp_class_nzero(x);
1650 /* Try finding the high byte, and assume it's highest bit
1651 * is the sign. This assumption is probably wrong somewhere. */
1652# elif defined(USE_LONG_DOUBLE) && LONG_DOUBLEKIND == LONG_DOUBLE_IS_X86_80_BIT_LITTLE_ENDIAN
1653 return (((unsigned char *)&x)[9] & 0x80);
1654# elif defined(NV_LITTLE_ENDIAN)
1655 /* Note that NVSIZE is sizeof(NV), which would make the below be
1656 * wrong if the end bytes are unused, which happens with the x86
1657 * 80-bit long doubles, which is why take care of that above. */
1658 return (((unsigned char *)&x)[NVSIZE - 1] & 0x80);
1659# elif defined(NV_BIG_ENDIAN)
1660 return (((unsigned char *)&x)[0] & 0x80);
1661# else
1662 /* This last resort fallback is wrong for the negative zero. */
1663 return (x < 0.0) ? 1 : 0;
1664# endif
1665}
1666#endif
1667
1668/*
1669 * ex: set ts=8 sts=4 sw=4 et:
1670 */