This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
Avoid unused warnings from locale-less systems.
[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 I<start> and I<*len> give the string to scan, I<*flags> gives
111conversion flags, and I<result> should be 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 I<*flags>, encountering an
114invalid character will also trigger a warning.
115On return I<*len> is set to the length of the scanned string,
116and I<*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 I<*result>. If the value is > UV_MAX C<grok_bin>
120returns UV_MAX, sets C<PERL_SCAN_GREATER_THAN_UV_MAX> in the output flags,
121and writes the value to I<*result> (or the value is discarded if I<result>
122is NULL).
123
124The binary number may optionally be prefixed with "0b" or "b" unless
125C<PERL_SCAN_DISALLOW_PREFIX> is set in I<*flags> on entry. If
126C<PERL_SCAN_ALLOW_UNDERSCORES> is set in I<*flags> then the binary
127number may use '_' 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 (s[0] == 'b' || s[0] == 'B') {
157 s++;
158 len--;
159 }
160 else if (len >= 2 && s[0] == '0' && (s[1] == 'b' || 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 I<start> and I<*len_p> give the string to scan, I<*flags> gives
234conversion flags, and I<result> should be 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 I<*flags>, encountering an
237invalid character will also trigger a warning.
238On return I<*len> is set to the length of the scanned string,
239and I<*flags> gives output flags.
240
241If the value is <= UV_MAX it is returned as a UV, the output flags are clear,
242and nothing is written to I<*result>. If the value is > UV_MAX C<grok_hex>
243returns UV_MAX, sets C<PERL_SCAN_GREATER_THAN_UV_MAX> in the output flags,
244and writes the value to I<*result> (or the value is discarded if I<result>
245is NULL).
246
247The hex number may optionally be prefixed with "0x" or "x" unless
248C<PERL_SCAN_DISALLOW_PREFIX> is set in I<*flags> on entry. If
249C<PERL_SCAN_ALLOW_UNDERSCORES> is set in I<*flags> then the hex
250number may use '_' 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 that are still 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 (s[0] == 'x' || s[0] == 'X') {
278 s++;
279 len--;
280 }
281 else if (len >= 2 && s[0] == '0' && (s[1] == 'x' || 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 I<start> and I<*len> give the string to scan, I<*flags> gives
355conversion flags, and I<result> should be 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 I<*flags>, encountering an
3588 or 9 will also trigger a warning.
359On return I<*len> is set to the length of the scanned string,
360and I<*flags> gives output flags.
361
362If the value is <= UV_MAX it is returned as a UV, the output flags are clear,
363and nothing is written to I<*result>. If the value is > UV_MAX C<grok_oct>
364returns UV_MAX, sets C<PERL_SCAN_GREATER_THAN_UV_MAX> in the output flags,
365and writes the value to I<*result> (or the value is discarded if I<result>
366is NULL).
367
368If C<PERL_SCAN_ALLOW_UNDERSCORES> is set in I<*flags> then the octal
369number may use '_' 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 DECLARE_STORE_LC_NUMERIC_SET_TO_NEEDED();
526 if (PL_numeric_radix_sv) {
527 STRLEN len;
528 const char * const radix = SvPV(PL_numeric_radix_sv, len);
529 if (*sp + len <= send && memEQ(*sp, radix, len)) {
530 *sp += len;
531 RESTORE_LC_NUMERIC();
532 return TRUE;
533 }
534 }
535 RESTORE_LC_NUMERIC();
536 }
537 /* always try "." if numeric radix didn't match because
538 * we may have data from different locales mixed */
539#endif
540
541 PERL_ARGS_ASSERT_GROK_NUMERIC_RADIX;
542
543 if (*sp < send && **sp == '.') {
544 ++*sp;
545 return TRUE;
546 }
547 return FALSE;
548}
549
550/*
551=for apidoc grok_number_flags
552
553Recognise (or not) a number. The type of the number is returned
554(0 if unrecognised), otherwise it is a bit-ORed combination of
555IS_NUMBER_IN_UV, IS_NUMBER_GREATER_THAN_UV_MAX, IS_NUMBER_NOT_INT,
556IS_NUMBER_NEG, IS_NUMBER_INFINITY, IS_NUMBER_NAN (defined in perl.h).
557
558If the value of the number can fit in a UV, it is returned in the *valuep
559IS_NUMBER_IN_UV will be set to indicate that *valuep is valid, IS_NUMBER_IN_UV
560will never be set unless *valuep is valid, but *valuep may have been assigned
561to during processing even though IS_NUMBER_IN_UV is not set on return.
562If valuep is NULL, IS_NUMBER_IN_UV will be set for the same cases as when
563valuep is non-NULL, but no actual assignment (or SEGV) will occur.
564
565IS_NUMBER_NOT_INT will be set with IS_NUMBER_IN_UV if trailing decimals were
566seen (in which case *valuep gives the true value truncated to an integer), and
567IS_NUMBER_NEG if the number is negative (in which case *valuep holds the
568absolute value). IS_NUMBER_IN_UV is not set if e notation was used or the
569number is larger than a UV.
570
571C<flags> allows only C<PERL_SCAN_TRAILING>, which allows for trailing
572non-numeric text on an otherwise successful I<grok>, setting
573C<IS_NUMBER_TRAILING> on the result.
574
575=for apidoc grok_number
576
577Identical to grok_number_flags() with flags set to zero.
578
579=cut
580 */
581int
582Perl_grok_number(pTHX_ const char *pv, STRLEN len, UV *valuep)
583{
584 PERL_ARGS_ASSERT_GROK_NUMBER;
585
586 return grok_number_flags(pv, len, valuep, 0);
587}
588
589int
590Perl_grok_number_flags(pTHX_ const char *pv, STRLEN len, UV *valuep, U32 flags)
591{
592 const char *s = pv;
593 const char * const send = pv + len;
594 const UV max_div_10 = UV_MAX / 10;
595 const char max_mod_10 = UV_MAX % 10;
596 int numtype = 0;
597 int sawinf = 0;
598 int sawnan = 0;
599
600 PERL_ARGS_ASSERT_GROK_NUMBER_FLAGS;
601
602 while (s < send && isSPACE(*s))
603 s++;
604 if (s == send) {
605 return 0;
606 } else if (*s == '-') {
607 s++;
608 numtype = IS_NUMBER_NEG;
609 }
610 else if (*s == '+')
611 s++;
612
613 if (s == send)
614 return 0;
615
616 /* next must be digit or the radix separator or beginning of infinity */
617 if (isDIGIT(*s)) {
618 /* UVs are at least 32 bits, so the first 9 decimal digits cannot
619 overflow. */
620 UV value = *s - '0';
621 /* This construction seems to be more optimiser friendly.
622 (without it gcc does the isDIGIT test and the *s - '0' separately)
623 With it gcc on arm is managing 6 instructions (6 cycles) per digit.
624 In theory the optimiser could deduce how far to unroll the loop
625 before checking for overflow. */
626 if (++s < send) {
627 int digit = *s - '0';
628 if (digit >= 0 && digit <= 9) {
629 value = value * 10 + digit;
630 if (++s < send) {
631 digit = *s - '0';
632 if (digit >= 0 && digit <= 9) {
633 value = value * 10 + digit;
634 if (++s < send) {
635 digit = *s - '0';
636 if (digit >= 0 && digit <= 9) {
637 value = value * 10 + digit;
638 if (++s < send) {
639 digit = *s - '0';
640 if (digit >= 0 && digit <= 9) {
641 value = value * 10 + digit;
642 if (++s < send) {
643 digit = *s - '0';
644 if (digit >= 0 && digit <= 9) {
645 value = value * 10 + digit;
646 if (++s < send) {
647 digit = *s - '0';
648 if (digit >= 0 && digit <= 9) {
649 value = value * 10 + digit;
650 if (++s < send) {
651 digit = *s - '0';
652 if (digit >= 0 && digit <= 9) {
653 value = value * 10 + digit;
654 if (++s < send) {
655 digit = *s - '0';
656 if (digit >= 0 && digit <= 9) {
657 value = value * 10 + digit;
658 if (++s < send) {
659 /* Now got 9 digits, so need to check
660 each time for overflow. */
661 digit = *s - '0';
662 while (digit >= 0 && digit <= 9
663 && (value < max_div_10
664 || (value == max_div_10
665 && digit <= max_mod_10))) {
666 value = value * 10 + digit;
667 if (++s < send)
668 digit = *s - '0';
669 else
670 break;
671 }
672 if (digit >= 0 && digit <= 9
673 && (s < send)) {
674 /* value overflowed.
675 skip the remaining digits, don't
676 worry about setting *valuep. */
677 do {
678 s++;
679 } while (s < send && isDIGIT(*s));
680 numtype |=
681 IS_NUMBER_GREATER_THAN_UV_MAX;
682 goto skip_value;
683 }
684 }
685 }
686 }
687 }
688 }
689 }
690 }
691 }
692 }
693 }
694 }
695 }
696 }
697 }
698 }
699 }
700 }
701 numtype |= IS_NUMBER_IN_UV;
702 if (valuep)
703 *valuep = value;
704
705 skip_value:
706 if (GROK_NUMERIC_RADIX(&s, send)) {
707 numtype |= IS_NUMBER_NOT_INT;
708 while (s < send && isDIGIT(*s)) /* optional digits after the radix */
709 s++;
710 }
711 }
712 else if (GROK_NUMERIC_RADIX(&s, send)) {
713 numtype |= IS_NUMBER_NOT_INT | IS_NUMBER_IN_UV; /* valuep assigned below */
714 /* no digits before the radix means we need digits after it */
715 if (s < send && isDIGIT(*s)) {
716 do {
717 s++;
718 } while (s < send && isDIGIT(*s));
719 if (valuep) {
720 /* integer approximation is valid - it's 0. */
721 *valuep = 0;
722 }
723 }
724 else
725 return 0;
726 } else if (*s == 'I' || *s == 'i') {
727 s++; if (s == send || (*s != 'N' && *s != 'n')) return 0;
728 s++; if (s == send || (*s != 'F' && *s != 'f')) return 0;
729 s++; if (s < send && (*s == 'I' || *s == 'i')) {
730 s++; if (s == send || (*s != 'N' && *s != 'n')) return 0;
731 s++; if (s == send || (*s != 'I' && *s != 'i')) return 0;
732 s++; if (s == send || (*s != 'T' && *s != 't')) return 0;
733 s++; if (s == send || (*s != 'Y' && *s != 'y')) return 0;
734 s++;
735 }
736 sawinf = 1;
737 } else if (*s == 'N' || *s == 'n') {
738 /* XXX TODO: There are signaling NaNs and quiet NaNs. */
739 s++; if (s == send || (*s != 'A' && *s != 'a')) return 0;
740 s++; if (s == send || (*s != 'N' && *s != 'n')) return 0;
741 s++;
742 sawnan = 1;
743 } else
744 return 0;
745
746 if (sawinf) {
747 numtype &= IS_NUMBER_NEG; /* Keep track of sign */
748 numtype |= IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT;
749 } else if (sawnan) {
750 numtype &= IS_NUMBER_NEG; /* Keep track of sign */
751 numtype |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
752 } else if (s < send) {
753 /* we can have an optional exponent part */
754 if (*s == 'e' || *s == 'E') {
755 s++;
756 if (s < send && (*s == '-' || *s == '+'))
757 s++;
758 if (s < send && isDIGIT(*s)) {
759 do {
760 s++;
761 } while (s < send && isDIGIT(*s));
762 }
763 else if (flags & PERL_SCAN_TRAILING)
764 return numtype | IS_NUMBER_TRAILING;
765 else
766 return 0;
767
768 /* The only flag we keep is sign. Blow away any "it's UV" */
769 numtype &= IS_NUMBER_NEG;
770 numtype |= IS_NUMBER_NOT_INT;
771 }
772 }
773 while (s < send && isSPACE(*s))
774 s++;
775 if (s >= send)
776 return numtype;
777 if (len == 10 && memEQ(pv, "0 but true", 10)) {
778 if (valuep)
779 *valuep = 0;
780 return IS_NUMBER_IN_UV;
781 }
782 else if (flags & PERL_SCAN_TRAILING) {
783 return numtype | IS_NUMBER_TRAILING;
784 }
785
786 return 0;
787}
788
789STATIC NV
790S_mulexp10(NV value, I32 exponent)
791{
792 NV result = 1.0;
793 NV power = 10.0;
794 bool negative = 0;
795 I32 bit;
796
797 if (exponent == 0)
798 return value;
799 if (value == 0)
800 return (NV)0;
801
802 /* On OpenVMS VAX we by default use the D_FLOAT double format,
803 * and that format does not have *easy* capabilities [1] for
804 * overflowing doubles 'silently' as IEEE fp does. We also need
805 * to support G_FLOAT on both VAX and Alpha, and though the exponent
806 * range is much larger than D_FLOAT it still doesn't do silent
807 * overflow. Therefore we need to detect early whether we would
808 * overflow (this is the behaviour of the native string-to-float
809 * conversion routines, and therefore of native applications, too).
810 *
811 * [1] Trying to establish a condition handler to trap floating point
812 * exceptions is not a good idea. */
813
814 /* In UNICOS and in certain Cray models (such as T90) there is no
815 * IEEE fp, and no way at all from C to catch fp overflows gracefully.
816 * There is something you can do if you are willing to use some
817 * inline assembler: the instruction is called DFI-- but that will
818 * disable *all* floating point interrupts, a little bit too large
819 * a hammer. Therefore we need to catch potential overflows before
820 * it's too late. */
821
822#if ((defined(VMS) && !defined(_IEEE_FP)) || defined(_UNICOS)) && defined(NV_MAX_10_EXP)
823 STMT_START {
824 const NV exp_v = log10(value);
825 if (exponent >= NV_MAX_10_EXP || exponent + exp_v >= NV_MAX_10_EXP)
826 return NV_MAX;
827 if (exponent < 0) {
828 if (-(exponent + exp_v) >= NV_MAX_10_EXP)
829 return 0.0;
830 while (-exponent >= NV_MAX_10_EXP) {
831 /* combination does not overflow, but 10^(-exponent) does */
832 value /= 10;
833 ++exponent;
834 }
835 }
836 } STMT_END;
837#endif
838
839 if (exponent < 0) {
840 negative = 1;
841 exponent = -exponent;
842#ifdef NV_MAX_10_EXP
843 /* for something like 1234 x 10^-309, the action of calculating
844 * the intermediate value 10^309 then returning 1234 / (10^309)
845 * will fail, since 10^309 becomes infinity. In this case try to
846 * refactor it as 123 / (10^308) etc.
847 */
848 while (value && exponent > NV_MAX_10_EXP) {
849 exponent--;
850 value /= 10;
851 }
852#endif
853 }
854 for (bit = 1; exponent; bit <<= 1) {
855 if (exponent & bit) {
856 exponent ^= bit;
857 result *= power;
858 /* Floating point exceptions are supposed to be turned off,
859 * but if we're obviously done, don't risk another iteration.
860 */
861 if (exponent == 0) break;
862 }
863 power *= power;
864 }
865 return negative ? value / result : value * result;
866}
867
868NV
869Perl_my_atof(pTHX_ const char* s)
870{
871 NV x = 0.0;
872#ifdef USE_LOCALE_NUMERIC
873 PERL_ARGS_ASSERT_MY_ATOF;
874
875 {
876 DECLARE_STORE_LC_NUMERIC_SET_TO_NEEDED();
877 if (PL_numeric_radix_sv && IN_LC(LC_NUMERIC)) {
878 const char *standard = NULL, *local = NULL;
879 bool use_standard_radix;
880
881 /* Look through the string for the first thing that looks like a
882 * decimal point: either the value in the current locale or the
883 * standard fallback of '.'. The one which appears earliest in the
884 * input string is the one that we should have atof look for. Note
885 * that we have to determine this beforehand because on some
886 * systems, Perl_atof2 is just a wrapper around the system's atof.
887 * */
888 standard = strchr(s, '.');
889 local = strstr(s, SvPV_nolen(PL_numeric_radix_sv));
890
891 use_standard_radix = standard && (!local || standard < local);
892
893 if (use_standard_radix)
894 SET_NUMERIC_STANDARD();
895
896 Perl_atof2(s, x);
897
898 if (use_standard_radix)
899 SET_NUMERIC_LOCAL();
900 }
901 else
902 Perl_atof2(s, x);
903 RESTORE_LC_NUMERIC();
904 }
905#else
906 Perl_atof2(s, x);
907#endif
908 return x;
909}
910
911char*
912Perl_my_atof2(pTHX_ const char* orig, NV* value)
913{
914 NV result[3] = {0.0, 0.0, 0.0};
915 const char* s = orig;
916#ifdef USE_PERL_ATOF
917 UV accumulator[2] = {0,0}; /* before/after dp */
918 bool negative = 0;
919 const char* send = s + strlen(orig) - 1;
920 bool seen_digit = 0;
921 I32 exp_adjust[2] = {0,0};
922 I32 exp_acc[2] = {-1, -1};
923 /* the current exponent adjust for the accumulators */
924 I32 exponent = 0;
925 I32 seen_dp = 0;
926 I32 digit = 0;
927 I32 old_digit = 0;
928 I32 sig_digits = 0; /* noof significant digits seen so far */
929
930 PERL_ARGS_ASSERT_MY_ATOF2;
931
932/* There is no point in processing more significant digits
933 * than the NV can hold. Note that NV_DIG is a lower-bound value,
934 * while we need an upper-bound value. We add 2 to account for this;
935 * since it will have been conservative on both the first and last digit.
936 * For example a 32-bit mantissa with an exponent of 4 would have
937 * exact values in the set
938 * 4
939 * 8
940 * ..
941 * 17179869172
942 * 17179869176
943 * 17179869180
944 *
945 * where for the purposes of calculating NV_DIG we would have to discount
946 * both the first and last digit, since neither can hold all values from
947 * 0..9; but for calculating the value we must examine those two digits.
948 */
949#ifdef MAX_SIG_DIG_PLUS
950 /* It is not necessarily the case that adding 2 to NV_DIG gets all the
951 possible digits in a NV, especially if NVs are not IEEE compliant
952 (e.g., long doubles on IRIX) - Allen <allens@cpan.org> */
953# define MAX_SIG_DIGITS (NV_DIG+MAX_SIG_DIG_PLUS)
954#else
955# define MAX_SIG_DIGITS (NV_DIG+2)
956#endif
957
958/* the max number we can accumulate in a UV, and still safely do 10*N+9 */
959#define MAX_ACCUMULATE ( (UV) ((UV_MAX - 9)/10))
960
961 /* leading whitespace */
962 while (isSPACE(*s))
963 ++s;
964
965 /* sign */
966 switch (*s) {
967 case '-':
968 negative = 1;
969 /* FALLTHROUGH */
970 case '+':
971 ++s;
972 }
973
974 /* punt to strtod for NaN/Inf; if no support for it there, tough luck */
975
976#ifdef HAS_STRTOD
977 if (*s == 'n' || *s == 'N' || *s == 'i' || *s == 'I') {
978 const char *p = negative ? s - 1 : s;
979 char *endp;
980 NV rslt;
981 rslt = strtod(p, &endp);
982 if (endp != p) {
983 *value = rslt;
984 return (char *)endp;
985 }
986 }
987#endif
988
989 /* we accumulate digits into an integer; when this becomes too
990 * large, we add the total to NV and start again */
991
992 while (1) {
993 if (isDIGIT(*s)) {
994 seen_digit = 1;
995 old_digit = digit;
996 digit = *s++ - '0';
997 if (seen_dp)
998 exp_adjust[1]++;
999
1000 /* don't start counting until we see the first significant
1001 * digit, eg the 5 in 0.00005... */
1002 if (!sig_digits && digit == 0)
1003 continue;
1004
1005 if (++sig_digits > MAX_SIG_DIGITS) {
1006 /* limits of precision reached */
1007 if (digit > 5) {
1008 ++accumulator[seen_dp];
1009 } else if (digit == 5) {
1010 if (old_digit % 2) { /* round to even - Allen */
1011 ++accumulator[seen_dp];
1012 }
1013 }
1014 if (seen_dp) {
1015 exp_adjust[1]--;
1016 } else {
1017 exp_adjust[0]++;
1018 }
1019 /* skip remaining digits */
1020 while (isDIGIT(*s)) {
1021 ++s;
1022 if (! seen_dp) {
1023 exp_adjust[0]++;
1024 }
1025 }
1026 /* warn of loss of precision? */
1027 }
1028 else {
1029 if (accumulator[seen_dp] > MAX_ACCUMULATE) {
1030 /* add accumulator to result and start again */
1031 result[seen_dp] = S_mulexp10(result[seen_dp],
1032 exp_acc[seen_dp])
1033 + (NV)accumulator[seen_dp];
1034 accumulator[seen_dp] = 0;
1035 exp_acc[seen_dp] = 0;
1036 }
1037 accumulator[seen_dp] = accumulator[seen_dp] * 10 + digit;
1038 ++exp_acc[seen_dp];
1039 }
1040 }
1041 else if (!seen_dp && GROK_NUMERIC_RADIX(&s, send)) {
1042 seen_dp = 1;
1043 if (sig_digits > MAX_SIG_DIGITS) {
1044 do {
1045 ++s;
1046 } while (isDIGIT(*s));
1047 break;
1048 }
1049 }
1050 else {
1051 break;
1052 }
1053 }
1054
1055 result[0] = S_mulexp10(result[0], exp_acc[0]) + (NV)accumulator[0];
1056 if (seen_dp) {
1057 result[1] = S_mulexp10(result[1], exp_acc[1]) + (NV)accumulator[1];
1058 }
1059
1060 if (seen_digit && (*s == 'e' || *s == 'E')) {
1061 bool expnegative = 0;
1062
1063 ++s;
1064 switch (*s) {
1065 case '-':
1066 expnegative = 1;
1067 /* FALLTHROUGH */
1068 case '+':
1069 ++s;
1070 }
1071 while (isDIGIT(*s))
1072 exponent = exponent * 10 + (*s++ - '0');
1073 if (expnegative)
1074 exponent = -exponent;
1075 }
1076
1077
1078
1079 /* now apply the exponent */
1080
1081 if (seen_dp) {
1082 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0])
1083 + S_mulexp10(result[1],exponent-exp_adjust[1]);
1084 } else {
1085 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0]);
1086 }
1087
1088 /* now apply the sign */
1089 if (negative)
1090 result[2] = -result[2];
1091#endif /* USE_PERL_ATOF */
1092 *value = result[2];
1093 return (char *)s;
1094}
1095
1096#if ! defined(HAS_MODFL) && defined(HAS_AINTL) && defined(HAS_COPYSIGNL)
1097long double
1098Perl_my_modfl(long double x, long double *ip)
1099{
1100 *ip = aintl(x);
1101 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
1102}
1103#endif
1104
1105#if ! defined(HAS_FREXPL) && defined(HAS_ILOGBL) && defined(HAS_SCALBNL)
1106long double
1107Perl_my_frexpl(long double x, int *e) {
1108 *e = x == 0.0L ? 0 : ilogbl(x) + 1;
1109 return (scalbnl(x, -*e));
1110}
1111#endif
1112
1113/*
1114=for apidoc Perl_signbit
1115
1116Return a non-zero integer if the sign bit on an NV is set, and 0 if
1117it is not.
1118
1119If Configure detects this system has a signbit() that will work with
1120our NVs, then we just use it via the #define in perl.h. Otherwise,
1121fall back on this implementation. As a first pass, this gets everything
1122right except -0.0. Alas, catching -0.0 is the main use for this function,
1123so this is not too helpful yet. Still, at least we have the scaffolding
1124in place to support other systems, should that prove useful.
1125
1126
1127Configure notes: This function is called 'Perl_signbit' instead of a
1128plain 'signbit' because it is easy to imagine a system having a signbit()
1129function or macro that doesn't happen to work with our particular choice
1130of NVs. We shouldn't just re-#define signbit as Perl_signbit and expect
1131the standard system headers to be happy. Also, this is a no-context
1132function (no pTHX_) because Perl_signbit() is usually re-#defined in
1133perl.h as a simple macro call to the system's signbit().
1134Users should just always call Perl_signbit().
1135
1136=cut
1137*/
1138#if !defined(HAS_SIGNBIT)
1139int
1140Perl_signbit(NV x) {
1141 return (x < 0.0) ? 1 : 0;
1142}
1143#endif
1144
1145/*
1146 * Local variables:
1147 * c-indentation-style: bsd
1148 * c-basic-offset: 4
1149 * indent-tabs-mode: nil
1150 * End:
1151 *
1152 * ex: set ts=8 sts=4 sw=4 et:
1153 */