This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
infnan: mention the unusual semantics of "numeric".
[perl5.git] / numeric.c
CommitLineData
98994639
HS
1/* numeric.c
2 *
663f364b 3 * Copyright (C) 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001,
1129b882 4 * 2002, 2003, 2004, 2005, 2006, 2007, 2008 by Larry Wall and others
98994639
HS
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/*
4ac71550
TC
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"]
98994639
HS
16 */
17
ccfc67b7
JH
18/*
19=head1 Numeric functions
166f8a29 20
7fefc6c1
KW
21=cut
22
166f8a29
DM
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
ccfc67b7
JH
26*/
27
98994639
HS
28#include "EXTERN.h"
29#define PERL_IN_NUMERIC_C
30#include "perl.h"
31
32U32
ddeaf645 33Perl_cast_ulong(NV f)
98994639
HS
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
ddeaf645 51Perl_cast_i32(NV f)
98994639
HS
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
ddeaf645 69Perl_cast_iv(NV f)
98994639
HS
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
ddeaf645 88Perl_cast_uv(NV f)
98994639
HS
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
53305cf1
NC
105/*
106=for apidoc grok_bin
98994639 107
53305cf1
NC
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.
7b667b5f
MHM
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.
53305cf1 117
7fc63493 118If the value is <= C<UV_MAX> it is returned as a UV, the output flags are clear,
72d33970 119and nothing is written to I<*result>. If the value is > UV_MAX C<grok_bin>
53305cf1
NC
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
7b667b5f 124The binary number may optionally be prefixed with "0b" or "b" unless
72d33970 125C<PERL_SCAN_DISALLOW_PREFIX> is set in I<*flags> on entry. If
a4c04bdc 126C<PERL_SCAN_ALLOW_UNDERSCORES> is set in I<*flags> then the binary
53305cf1
NC
127number may use '_' characters to separate digits.
128
129=cut
02470786
KW
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.
53305cf1
NC
134 */
135
136UV
7918f24d
NC
137Perl_grok_bin(pTHX_ const char *start, STRLEN *len_p, I32 *flags, NV *result)
138{
53305cf1
NC
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;
f2338a2e 145 const bool allow_underscores = cBOOL(*flags & PERL_SCAN_ALLOW_UNDERSCORES);
53305cf1 146 bool overflowed = FALSE;
7fc63493 147 char bit;
53305cf1 148
7918f24d
NC
149 PERL_ARGS_ASSERT_GROK_BIN;
150
a4c04bdc
NC
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) {
305b8651 156 if (isALPHA_FOLD_EQ(s[0], 'b')) {
a4c04bdc
NC
157 s++;
158 len--;
159 }
305b8651 160 else if (len >= 2 && s[0] == '0' && (isALPHA_FOLD_EQ(s[1], 'b'))) {
a4c04bdc
NC
161 s+=2;
162 len-=2;
163 }
164 }
53305cf1
NC
165 }
166
7fc63493 167 for (; len-- && (bit = *s); s++) {
53305cf1
NC
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. */
dcbac5bb 179 /* diag_listed_as: Integer overflow in %s number */
9b387841
NC
180 Perl_ck_warner_d(aTHX_ packWARN(WARN_OVERFLOW),
181 "Integer overflow in binary number");
53305cf1
NC
182 overflowed = TRUE;
183 value_nv = (NV) value;
184 }
185 value_nv *= 2.0;
98994639 186 /* If an NV has not enough bits in its mantissa to
d1be9408 187 * represent a UV this summing of small low-order numbers
98994639
HS
188 * is a waste of time (because the NV cannot preserve
189 * the low-order bits anyway): we could just remember when
53305cf1 190 * did we overflow and in the end just multiply value_nv by the
98994639 191 * right amount. */
53305cf1
NC
192 value_nv += (NV)(bit - '0');
193 continue;
194 }
195 if (bit == '_' && len && allow_underscores && (bit = s[1])
196 && (bit == '0' || bit == '1'))
98994639
HS
197 {
198 --len;
199 ++s;
53305cf1 200 goto redo;
98994639 201 }
a2a5de95
NC
202 if (!(*flags & PERL_SCAN_SILENT_ILLDIGIT))
203 Perl_ck_warner(aTHX_ packWARN(WARN_DIGIT),
204 "Illegal binary digit '%c' ignored", *s);
53305cf1 205 break;
98994639 206 }
53305cf1
NC
207
208 if ( ( overflowed && value_nv > 4294967295.0)
98994639 209#if UVSIZE > 4
02470786
KW
210 || (!overflowed && value > 0xffffffff
211 && ! (*flags & PERL_SCAN_SILENT_NON_PORTABLE))
98994639
HS
212#endif
213 ) {
a2a5de95
NC
214 Perl_ck_warner(aTHX_ packWARN(WARN_PORTABLE),
215 "Binary number > 0b11111111111111111111111111111111 non-portable");
53305cf1
NC
216 }
217 *len_p = s - start;
218 if (!overflowed) {
219 *flags = 0;
220 return value;
98994639 221 }
53305cf1
NC
222 *flags = PERL_SCAN_GREATER_THAN_UV_MAX;
223 if (result)
224 *result = value_nv;
225 return UV_MAX;
98994639
HS
226}
227
53305cf1
NC
228/*
229=for apidoc grok_hex
230
231converts a string representing a hex number to numeric form.
232
c2da02fc 233On entry I<start> and I<*len_p> give the string to scan, I<*flags> gives
53305cf1 234conversion flags, and I<result> should be NULL or a pointer to an NV.
7b667b5f
MHM
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.
53305cf1
NC
240
241If the value is <= UV_MAX it is returned as a UV, the output flags are clear,
72d33970 242and nothing is written to I<*result>. If the value is > UV_MAX C<grok_hex>
53305cf1
NC
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
d1be9408 247The hex number may optionally be prefixed with "0x" or "x" unless
72d33970 248C<PERL_SCAN_DISALLOW_PREFIX> is set in I<*flags> on entry. If
a4c04bdc 249C<PERL_SCAN_ALLOW_UNDERSCORES> is set in I<*flags> then the hex
53305cf1
NC
250number may use '_' characters to separate digits.
251
252=cut
02470786
KW
253
254Not documented yet because experimental is C<PERL_SCAN_SILENT_NON_PORTABLE
baf48926 255which suppresses any message for non-portable numbers, but which are valid
02470786 256on this platform.
53305cf1
NC
257 */
258
259UV
7918f24d
NC
260Perl_grok_hex(pTHX_ const char *start, STRLEN *len_p, I32 *flags, NV *result)
261{
53305cf1
NC
262 const char *s = start;
263 STRLEN len = *len_p;
264 UV value = 0;
265 NV value_nv = 0;
53305cf1 266 const UV max_div_16 = UV_MAX / 16;
f2338a2e 267 const bool allow_underscores = cBOOL(*flags & PERL_SCAN_ALLOW_UNDERSCORES);
53305cf1 268 bool overflowed = FALSE;
98994639 269
7918f24d
NC
270 PERL_ARGS_ASSERT_GROK_HEX;
271
a4c04bdc
NC
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) {
305b8651 277 if (isALPHA_FOLD_EQ(s[0], 'x')) {
a4c04bdc
NC
278 s++;
279 len--;
280 }
305b8651 281 else if (len >= 2 && s[0] == '0' && (isALPHA_FOLD_EQ(s[1], 'x'))) {
a4c04bdc
NC
282 s+=2;
283 len-=2;
284 }
285 }
98994639
HS
286 }
287
288 for (; len-- && *s; s++) {
626ef089 289 if (isXDIGIT(*s)) {
53305cf1
NC
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) {
626ef089 296 value = (value << 4) | XDIGIT_VALUE(*s);
53305cf1
NC
297 continue;
298 }
299 /* Bah. We're just overflowed. */
dcbac5bb 300 /* diag_listed_as: Integer overflow in %s number */
9b387841
NC
301 Perl_ck_warner_d(aTHX_ packWARN(WARN_OVERFLOW),
302 "Integer overflow in hexadecimal number");
53305cf1
NC
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
d1be9408 308 * represent a UV this summing of small low-order numbers
53305cf1
NC
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. */
626ef089 313 value_nv += (NV) XDIGIT_VALUE(*s);
53305cf1
NC
314 continue;
315 }
316 if (*s == '_' && len && allow_underscores && s[1]
626ef089 317 && isXDIGIT(s[1]))
98994639
HS
318 {
319 --len;
320 ++s;
53305cf1 321 goto redo;
98994639 322 }
a2a5de95
NC
323 if (!(*flags & PERL_SCAN_SILENT_ILLDIGIT))
324 Perl_ck_warner(aTHX_ packWARN(WARN_DIGIT),
53305cf1
NC
325 "Illegal hexadecimal digit '%c' ignored", *s);
326 break;
327 }
328
329 if ( ( overflowed && value_nv > 4294967295.0)
330#if UVSIZE > 4
02470786
KW
331 || (!overflowed && value > 0xffffffff
332 && ! (*flags & PERL_SCAN_SILENT_NON_PORTABLE))
53305cf1
NC
333#endif
334 ) {
a2a5de95
NC
335 Perl_ck_warner(aTHX_ packWARN(WARN_PORTABLE),
336 "Hexadecimal number > 0xffffffff non-portable");
53305cf1
NC
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
7b667b5f
MHM
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
154bd527 3588 or 9 will also trigger a warning.
7b667b5f
MHM
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,
72d33970 363and nothing is written to I<*result>. If the value is > UV_MAX C<grok_oct>
7b667b5f
MHM
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.
53305cf1
NC
370
371=cut
02470786 372
333ae27c
KW
373Not documented yet because experimental is C<PERL_SCAN_SILENT_NON_PORTABLE>
374which suppresses any message for non-portable numbers, but which are valid
02470786 375on this platform.
53305cf1
NC
376 */
377
378UV
7918f24d
NC
379Perl_grok_oct(pTHX_ const char *start, STRLEN *len_p, I32 *flags, NV *result)
380{
53305cf1
NC
381 const char *s = start;
382 STRLEN len = *len_p;
383 UV value = 0;
384 NV value_nv = 0;
53305cf1 385 const UV max_div_8 = UV_MAX / 8;
f2338a2e 386 const bool allow_underscores = cBOOL(*flags & PERL_SCAN_ALLOW_UNDERSCORES);
53305cf1
NC
387 bool overflowed = FALSE;
388
7918f24d
NC
389 PERL_ARGS_ASSERT_GROK_OCT;
390
53305cf1 391 for (; len-- && *s; s++) {
626ef089 392 if (isOCTAL(*s)) {
53305cf1
NC
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) {
626ef089 399 value = (value << 3) | OCTAL_VALUE(*s);
53305cf1
NC
400 continue;
401 }
402 /* Bah. We're just overflowed. */
dcbac5bb 403 /* diag_listed_as: Integer overflow in %s number */
9b387841
NC
404 Perl_ck_warner_d(aTHX_ packWARN(WARN_OVERFLOW),
405 "Integer overflow in octal number");
53305cf1
NC
406 overflowed = TRUE;
407 value_nv = (NV) value;
408 }
409 value_nv *= 8.0;
98994639 410 /* If an NV has not enough bits in its mantissa to
d1be9408 411 * represent a UV this summing of small low-order numbers
98994639
HS
412 * is a waste of time (because the NV cannot preserve
413 * the low-order bits anyway): we could just remember when
53305cf1
NC
414 * did we overflow and in the end just multiply value_nv by the
415 * right amount of 8-tuples. */
626ef089 416 value_nv += (NV) OCTAL_VALUE(*s);
53305cf1
NC
417 continue;
418 }
626ef089
KW
419 if (*s == '_' && len && allow_underscores && isOCTAL(s[1])) {
420 --len;
421 ++s;
422 goto redo;
423 }
53305cf1 424 /* Allow \octal to work the DWIM way (that is, stop scanning
7b667b5f 425 * as soon as non-octal characters are seen, complain only if
626ef089
KW
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)) {
a2a5de95
NC
429 if (!(*flags & PERL_SCAN_SILENT_ILLDIGIT))
430 Perl_ck_warner(aTHX_ packWARN(WARN_DIGIT),
431 "Illegal octal digit '%c' ignored", *s);
53305cf1
NC
432 }
433 break;
98994639 434 }
53305cf1
NC
435
436 if ( ( overflowed && value_nv > 4294967295.0)
98994639 437#if UVSIZE > 4
02470786
KW
438 || (!overflowed && value > 0xffffffff
439 && ! (*flags & PERL_SCAN_SILENT_NON_PORTABLE))
98994639
HS
440#endif
441 ) {
a2a5de95
NC
442 Perl_ck_warner(aTHX_ packWARN(WARN_PORTABLE),
443 "Octal number > 037777777777 non-portable");
53305cf1
NC
444 }
445 *len_p = s - start;
446 if (!overflowed) {
447 *flags = 0;
448 return value;
98994639 449 }
53305cf1
NC
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
72d33970 459For backwards compatibility. Use C<grok_bin> instead.
53305cf1
NC
460
461=for apidoc scan_hex
462
72d33970 463For backwards compatibility. Use C<grok_hex> instead.
53305cf1
NC
464
465=for apidoc scan_oct
466
72d33970 467For backwards compatibility. Use C<grok_oct> instead.
53305cf1
NC
468
469=cut
470 */
471
472NV
73d840c0 473Perl_scan_bin(pTHX_ const char *start, STRLEN len, STRLEN *retlen)
53305cf1
NC
474{
475 NV rnv;
476 I32 flags = *retlen ? PERL_SCAN_ALLOW_UNDERSCORES : 0;
73d840c0 477 const UV ruv = grok_bin (start, &len, &flags, &rnv);
53305cf1 478
7918f24d
NC
479 PERL_ARGS_ASSERT_SCAN_BIN;
480
53305cf1
NC
481 *retlen = len;
482 return (flags & PERL_SCAN_GREATER_THAN_UV_MAX) ? rnv : (NV)ruv;
483}
484
485NV
73d840c0 486Perl_scan_oct(pTHX_ const char *start, STRLEN len, STRLEN *retlen)
53305cf1
NC
487{
488 NV rnv;
489 I32 flags = *retlen ? PERL_SCAN_ALLOW_UNDERSCORES : 0;
73d840c0 490 const UV ruv = grok_oct (start, &len, &flags, &rnv);
53305cf1 491
7918f24d
NC
492 PERL_ARGS_ASSERT_SCAN_OCT;
493
53305cf1
NC
494 *retlen = len;
495 return (flags & PERL_SCAN_GREATER_THAN_UV_MAX) ? rnv : (NV)ruv;
496}
497
498NV
73d840c0 499Perl_scan_hex(pTHX_ const char *start, STRLEN len, STRLEN *retlen)
53305cf1
NC
500{
501 NV rnv;
502 I32 flags = *retlen ? PERL_SCAN_ALLOW_UNDERSCORES : 0;
73d840c0 503 const UV ruv = grok_hex (start, &len, &flags, &rnv);
53305cf1 504
7918f24d
NC
505 PERL_ARGS_ASSERT_SCAN_HEX;
506
53305cf1
NC
507 *retlen = len;
508 return (flags & PERL_SCAN_GREATER_THAN_UV_MAX) ? rnv : (NV)ruv;
98994639
HS
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
7918f24d
NC
522 PERL_ARGS_ASSERT_GROK_NUMERIC_RADIX;
523
d6ded950 524 if (IN_LC(LC_NUMERIC)) {
21431899
KW
525 DECLARE_STORE_LC_NUMERIC_SET_TO_NEEDED();
526 if (PL_numeric_radix_sv) {
c5a7e38e
KW
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 }
21431899
KW
534 }
535 RESTORE_LC_NUMERIC();
98994639
HS
536 }
537 /* always try "." if numeric radix didn't match because
538 * we may have data from different locales mixed */
539#endif
7918f24d
NC
540
541 PERL_ARGS_ASSERT_GROK_NUMERIC_RADIX;
542
98994639
HS
543 if (*sp < send && **sp == '.') {
544 ++*sp;
545 return TRUE;
546 }
547 return FALSE;
548}
549
550/*
3f7602fa 551=for apidoc grok_number_flags
98994639
HS
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,
aa8b85de 556IS_NUMBER_NEG, IS_NUMBER_INFINITY, IS_NUMBER_NAN (defined in perl.h).
60939fb8 557
cd164854 558If the value of the number can fit in a UV, it is returned in the *valuep
60939fb8
NC
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.
98994639 570
3f7602fa
TC
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
98994639
HS
579=cut
580 */
581int
582Perl_grok_number(pTHX_ const char *pv, STRLEN len, UV *valuep)
583{
3f7602fa
TC
584 PERL_ARGS_ASSERT_GROK_NUMBER;
585
586 return grok_number_flags(pv, len, valuep, 0);
587}
588
ff4eb398
JH
589/*
590=for apidoc grok_infnan
591
592Helper for grok_number(), accepts various ways of spelling "infinity"
593or "not a number", and returns one of the following flag combinations:
594
595 IS_NUMBER_INFINITE
596 IS_NUMBER_NAN
597 IS_NUMBER_INFINITE | IS_NUMBER_NEG
598 IS_NUMBER_NAN | IS_NUMBER_NEG
599 0
600
b489e20f
JH
601possibly with IS_NUMBER_TRAILING.
602
ff4eb398
JH
603If an infinity or not-a-number is recognized, the *sp will point to
604one past the end of the recognized string. If the recognition fails,
605zero is returned, and the *sp will not move.
606
607=cut
608*/
609
610int
611Perl_grok_infnan(const char** sp, const char* send)
612{
613 const char* s = *sp;
614 int flags = 0;
e855f543 615 bool odh = FALSE; /* one dot hash: 1.#INF */
ff4eb398
JH
616
617 PERL_ARGS_ASSERT_GROK_INFNAN;
618
8c12dc63
JH
619 if (*s == '+') {
620 s++; if (s == send) return 0;
621 }
622 else if (*s == '-') {
ff4eb398
JH
623 flags |= IS_NUMBER_NEG; /* Yes, -NaN happens. Incorrect but happens. */
624 s++; if (s == send) return 0;
625 }
626
627 if (*s == '1') {
628 /* Visual C: 1.#SNAN, -1.#QNAN, 1#INF, 1#.IND (maybe also 1.#NAN) */
629 s++; if (s == send) return 0;
630 if (*s == '.') {
631 s++; if (s == send) return 0;
632 }
633 if (*s == '#') {
634 s++; if (s == send) return 0;
635 } else
636 return 0;
e855f543 637 odh = TRUE;
ff4eb398
JH
638 }
639
305b8651 640 if (isALPHA_FOLD_EQ(*s, 'I')) {
ff4eb398 641 /* INF or IND (1.#IND is indeterminate, a certain type of NAN) */
305b8651 642 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
ff4eb398 643 s++; if (s == send) return 0;
305b8651 644 if (isALPHA_FOLD_EQ(*s, 'F')) {
ff4eb398 645 s++;
3396ed30
JH
646 while (*s == '0') { /* 1.#INF00 */
647 s++;
648 }
b489e20f
JH
649 while (s < send && isSPACE(*s))
650 s++;
651 if (s < send && *s) {
3396ed30 652 flags |= IS_NUMBER_TRAILING;
fae4db12 653 }
ff4eb398
JH
654 flags |= IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT;
655 }
e855f543 656 else if (isALPHA_FOLD_EQ(*s, 'D') && odh) { /* 1.#IND */
ff4eb398
JH
657 s++;
658 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
fae4db12
JH
659 while (*s == '0') { /* 1.#IND00 */
660 s++;
661 }
1e9aa12f
JH
662 if (*s) {
663 flags |= IS_NUMBER_TRAILING;
664 }
ff4eb398
JH
665 } else
666 return 0;
ff4eb398
JH
667 }
668 else {
669 /* NAN */
305b8651 670 if (isALPHA_FOLD_EQ(*s, 'S') || isALPHA_FOLD_EQ(*s, 'Q')) {
ff4eb398
JH
671 /* snan, qNaN */
672 /* XXX do something with the snan/qnan difference */
ae776a2c 673 s++; if (s == send) return 0;
ff4eb398
JH
674 }
675
305b8651
KW
676 if (isALPHA_FOLD_EQ(*s, 'N')) {
677 s++; if (s == send || isALPHA_FOLD_NE(*s, 'A')) return 0;
678 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
ff4eb398
JH
679 s++;
680
ae776a2c
JH
681 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
682
8c12dc63
JH
683 /* NaN can be followed by various stuff (NaNQ, NaNS), but
684 * there are also multiple different NaN values, and some
685 * implementations output the "payload" values,
1e9aa12f 686 * e.g. NaN123, NAN(abc), while some legacy implementations
8c12dc63 687 * have weird stuff like NaN%. */
1e9aa12f
JH
688 if (isALPHA_FOLD_EQ(*s, 'q') ||
689 isALPHA_FOLD_EQ(*s, 's')) {
690 /* "nanq" or "nans" are ok, though generating
691 * these portably is tricky. */
692 s++;
693 }
694 if (*s == '(') {
695 /* C99 style "nan(123)" or Perlish equivalent "nan($uv)". */
696 const char *t;
697 UV nanval;
698 s++;
699 if (s == send) {
700 return flags | IS_NUMBER_TRAILING;
701 }
702 t = s + 1;
703 while (t < send && *t && *t != ')') {
704 t++;
705 }
706 if (t == send) {
707 return flags | IS_NUMBER_TRAILING;
708 }
709 if (*t == ')') {
710 int nantype =
711 grok_number_flags(s, t - s, &nanval,
712 PERL_SCAN_TRAILING |
713 PERL_SCAN_ALLOW_UNDERSCORES);
714 /* nanval result currently unused */
715 if ((nantype & IS_NUMBER_NOT_INT) ||
716 !(nantype && IS_NUMBER_IN_UV)) {
717 /* Certain configuration combinations where
718 * NVSIZE is greater than UVSIZE mean that a
719 * single UV cannot contain all the possible
720 * NaN payload bits. There would need to be
721 * some more generic syntax than "nan($uv)".
722 * Issues to keep in mind:
723 * (1) In most common cases there would
724 * not be an integral number of bytes that
725 * could be set, only a certain number of bits.
726 * For example for NVSIZE==UVSIZE there can be
727 * up to 52 bits in the payload, but one bit is
728 * commonly reserved for the signal/quiet bit,
729 * so 51 bits.
730 * (2) Endianness of the payload bits. If the
731 * payload is specified as an UV, the low-order
732 * bits of the UV are naturally little-endianed
733 * (rightmost) bits of the payload. */
734 return 0;
735 }
736 /* Unfortunately the grok_ interfaces don't tell
737 * the count of the consumed bytes, so we cannot
738 * figure out where the scanning left off.
739 * So we need to duplicate the basics of
740 * the scan ourselves. */
741 if (s[0] == '0' && s < t &&
742 isALPHA_FOLD_EQ(s[1], 'x')) {
743 const char *u = s + 2;
744 if (isXDIGIT(*u)) {
745 while (u < t &&
746 (isXDIGIT(*u) || *u == '_')) {
747 u++;
748 }
749 }
750 s = u;
751 } else if (isDIGIT(*s) && s < t) {
752 const char *u = s + 2;
753 while (u < t &&
754 (isDIGIT(*u) || *u == '_')) {
755 u++;
756 }
757 s = u;
758 }
759 /* XXX 0b... maybe, octal (really?) */
760 if (s < t) {
761 flags |= IS_NUMBER_TRAILING;
762 }
763 } else {
764 /* Looked like nan(...), but no close paren. */
765 flags |= IS_NUMBER_TRAILING;
766 }
b489e20f
JH
767 } else {
768 while (s < send && isSPACE(*s))
769 s++;
770 if (s < send && *s) {
771 /* Note that we here implicitly accept (parse as
772 * "nan", but with warnings) also any other weird
773 * trailing stuff for "nan". In the above we just
774 * check that if we got the C99-style "nan(...)",
775 * the "..." looks sane.
776 * If in future we accept more ways of specifying
777 * the nan payload, the accepting would happen around
778 * here. */
779 flags |= IS_NUMBER_TRAILING;
780 }
1e9aa12f 781 }
ae776a2c 782 s = send;
ff4eb398 783 }
ae776a2c
JH
784 else
785 return 0;
ff4eb398
JH
786 }
787
b489e20f
JH
788 while (s < send && isSPACE(*s))
789 s++;
790
a1fe7cea
JH
791 *sp = s;
792 return flags;
ff4eb398
JH
793}
794
945b524a
JH
795static const UV uv_max_div_10 = UV_MAX / 10;
796static const U8 uv_max_mod_10 = UV_MAX % 10;
797
3f7602fa
TC
798int
799Perl_grok_number_flags(pTHX_ const char *pv, STRLEN len, UV *valuep, U32 flags)
800{
60939fb8 801 const char *s = pv;
c4420975 802 const char * const send = pv + len;
ae776a2c 803 const char *d;
60939fb8 804 int numtype = 0;
60939fb8 805
3f7602fa 806 PERL_ARGS_ASSERT_GROK_NUMBER_FLAGS;
7918f24d 807
60939fb8
NC
808 while (s < send && isSPACE(*s))
809 s++;
810 if (s == send) {
811 return 0;
812 } else if (*s == '-') {
813 s++;
814 numtype = IS_NUMBER_NEG;
815 }
816 else if (*s == '+')
aa42a541 817 s++;
60939fb8
NC
818
819 if (s == send)
820 return 0;
821
ae776a2c 822 /* The first digit (after optional sign): note that might
8c12dc63 823 * also point to "infinity" or "nan", or "1.#INF". */
ae776a2c
JH
824 d = s;
825
8c12dc63 826 /* next must be digit or the radix separator or beginning of infinity/nan */
60939fb8
NC
827 if (isDIGIT(*s)) {
828 /* UVs are at least 32 bits, so the first 9 decimal digits cannot
829 overflow. */
830 UV value = *s - '0';
831 /* This construction seems to be more optimiser friendly.
832 (without it gcc does the isDIGIT test and the *s - '0' separately)
833 With it gcc on arm is managing 6 instructions (6 cycles) per digit.
834 In theory the optimiser could deduce how far to unroll the loop
835 before checking for overflow. */
58bb9ec3
NC
836 if (++s < send) {
837 int digit = *s - '0';
60939fb8
NC
838 if (digit >= 0 && digit <= 9) {
839 value = value * 10 + digit;
58bb9ec3
NC
840 if (++s < send) {
841 digit = *s - '0';
60939fb8
NC
842 if (digit >= 0 && digit <= 9) {
843 value = value * 10 + digit;
58bb9ec3
NC
844 if (++s < send) {
845 digit = *s - '0';
60939fb8
NC
846 if (digit >= 0 && digit <= 9) {
847 value = value * 10 + digit;
58bb9ec3
NC
848 if (++s < send) {
849 digit = *s - '0';
60939fb8
NC
850 if (digit >= 0 && digit <= 9) {
851 value = value * 10 + digit;
58bb9ec3
NC
852 if (++s < send) {
853 digit = *s - '0';
60939fb8
NC
854 if (digit >= 0 && digit <= 9) {
855 value = value * 10 + digit;
58bb9ec3
NC
856 if (++s < send) {
857 digit = *s - '0';
60939fb8
NC
858 if (digit >= 0 && digit <= 9) {
859 value = value * 10 + digit;
58bb9ec3
NC
860 if (++s < send) {
861 digit = *s - '0';
60939fb8
NC
862 if (digit >= 0 && digit <= 9) {
863 value = value * 10 + digit;
58bb9ec3
NC
864 if (++s < send) {
865 digit = *s - '0';
60939fb8
NC
866 if (digit >= 0 && digit <= 9) {
867 value = value * 10 + digit;
58bb9ec3 868 if (++s < send) {
60939fb8
NC
869 /* Now got 9 digits, so need to check
870 each time for overflow. */
58bb9ec3 871 digit = *s - '0';
60939fb8 872 while (digit >= 0 && digit <= 9
945b524a
JH
873 && (value < uv_max_div_10
874 || (value == uv_max_div_10
875 && digit <= uv_max_mod_10))) {
60939fb8 876 value = value * 10 + digit;
58bb9ec3
NC
877 if (++s < send)
878 digit = *s - '0';
60939fb8
NC
879 else
880 break;
881 }
882 if (digit >= 0 && digit <= 9
51bd16da 883 && (s < send)) {
60939fb8
NC
884 /* value overflowed.
885 skip the remaining digits, don't
886 worry about setting *valuep. */
887 do {
888 s++;
889 } while (s < send && isDIGIT(*s));
890 numtype |=
891 IS_NUMBER_GREATER_THAN_UV_MAX;
892 goto skip_value;
893 }
894 }
895 }
98994639 896 }
60939fb8
NC
897 }
898 }
899 }
900 }
901 }
902 }
903 }
904 }
905 }
906 }
907 }
98994639 908 }
60939fb8 909 }
98994639 910 }
60939fb8
NC
911 numtype |= IS_NUMBER_IN_UV;
912 if (valuep)
913 *valuep = value;
914
915 skip_value:
916 if (GROK_NUMERIC_RADIX(&s, send)) {
917 numtype |= IS_NUMBER_NOT_INT;
918 while (s < send && isDIGIT(*s)) /* optional digits after the radix */
919 s++;
98994639 920 }
60939fb8
NC
921 }
922 else if (GROK_NUMERIC_RADIX(&s, send)) {
923 numtype |= IS_NUMBER_NOT_INT | IS_NUMBER_IN_UV; /* valuep assigned below */
924 /* no digits before the radix means we need digits after it */
925 if (s < send && isDIGIT(*s)) {
926 do {
927 s++;
928 } while (s < send && isDIGIT(*s));
929 if (valuep) {
930 /* integer approximation is valid - it's 0. */
931 *valuep = 0;
932 }
98994639 933 }
60939fb8 934 else
ae776a2c 935 return 0;
ff4eb398 936 }
60939fb8 937
926f5fc6 938 if (s > d && s < send) {
60939fb8 939 /* we can have an optional exponent part */
305b8651 940 if (isALPHA_FOLD_EQ(*s, 'e')) {
60939fb8
NC
941 s++;
942 if (s < send && (*s == '-' || *s == '+'))
943 s++;
944 if (s < send && isDIGIT(*s)) {
945 do {
946 s++;
947 } while (s < send && isDIGIT(*s));
948 }
3f7602fa
TC
949 else if (flags & PERL_SCAN_TRAILING)
950 return numtype | IS_NUMBER_TRAILING;
60939fb8 951 else
3f7602fa
TC
952 return 0;
953
954 /* The only flag we keep is sign. Blow away any "it's UV" */
955 numtype &= IS_NUMBER_NEG;
956 numtype |= IS_NUMBER_NOT_INT;
60939fb8
NC
957 }
958 }
959 while (s < send && isSPACE(*s))
960 s++;
961 if (s >= send)
aa8b85de 962 return numtype;
60939fb8
NC
963 if (len == 10 && memEQ(pv, "0 but true", 10)) {
964 if (valuep)
965 *valuep = 0;
966 return IS_NUMBER_IN_UV;
967 }
8c12dc63
JH
968 /* We could be e.g. at "Inf" or "NaN", or at the "#" of "1.#INF". */
969 if ((s + 2 < send) && strchr("inqs#", toFOLD(*s))) {
970 /* Really detect inf/nan. Start at d, not s, since the above
971 * code might have already consumed the "1." or "1". */
972 int infnan = Perl_grok_infnan(&d, send);
973 if ((infnan & IS_NUMBER_INFINITY)) {
974 return (numtype | infnan); /* Keep sign for infinity. */
975 }
976 else if ((infnan & IS_NUMBER_NAN)) {
977 return (numtype | infnan) & ~IS_NUMBER_NEG; /* Clear sign for nan. */
978 }
979 }
3f7602fa
TC
980 else if (flags & PERL_SCAN_TRAILING) {
981 return numtype | IS_NUMBER_TRAILING;
982 }
983
60939fb8 984 return 0;
98994639
HS
985}
986
6313e544 987/*
d62b8c6a 988=for apidoc grok_atou
6313e544 989
d62b8c6a 990grok_atou is a safer replacement for atoi and strtol.
6313e544 991
d62b8c6a
JH
992grok_atou parses a C-style zero-byte terminated string, looking for
993a decimal unsigned integer.
338aa8b0 994
d62b8c6a
JH
995Returns the unsigned integer, if a valid value can be parsed
996from the beginning of the string.
f4379102 997
d62b8c6a 998Accepts only the decimal digits '0'..'9'.
6313e544 999
d62b8c6a
JH
1000As opposed to atoi or strtol, grok_atou does NOT allow optional
1001leading whitespace, or negative inputs. If such features are
1002required, the calling code needs to explicitly implement those.
6313e544 1003
d62b8c6a 1004If a valid value cannot be parsed, returns either zero (if non-digits
75feedba 1005are met before any digits) or UV_MAX (if the value overflows).
6313e544 1006
d62b8c6a
JH
1007Note that extraneous leading zeros also count as an overflow
1008(meaning that only "0" is the zero).
338aa8b0 1009
d62b8c6a 1010On failure, the *endptr is also set to NULL, unless endptr is NULL.
338aa8b0
JH
1011
1012Trailing non-digit bytes are allowed if the endptr is non-NULL.
6313e544
JH
1013On return the *endptr will contain the pointer to the first non-digit byte.
1014
6313e544 1015If the endptr is NULL, the first non-digit byte MUST be
f4379102 1016the zero byte terminating the pv, or zero will be returned.
6313e544 1017
d62b8c6a
JH
1018Background: atoi has severe problems with illegal inputs, it cannot be
1019used for incremental parsing, and therefore should be avoided
1020atoi and strtol are also affected by locale settings, which can also be
1021seen as a bug (global state controlled by user environment).
1022
6313e544
JH
1023=cut
1024*/
1025
75feedba 1026UV
6313e544
JH
1027Perl_grok_atou(const char *pv, const char** endptr)
1028{
1029 const char* s = pv;
1030 const char** eptr;
1031 const char* end2; /* Used in case endptr is NULL. */
75feedba 1032 UV val = 0; /* The return value. */
6313e544
JH
1033
1034 PERL_ARGS_ASSERT_GROK_ATOU;
1035
1036 eptr = endptr ? endptr : &end2;
75feedba
JH
1037 if (isDIGIT(*s)) {
1038 /* Single-digit inputs are quite common. */
6313e544 1039 val = *s++ - '0';
75feedba
JH
1040 if (isDIGIT(*s)) {
1041 /* Extra leading zeros cause overflow. */
1042 if (val == 0) {
1043 *eptr = NULL;
1044 return UV_MAX;
1045 }
1046 while (isDIGIT(*s)) {
1047 /* This could be unrolled like in grok_number(), but
1048 * the expected uses of this are not speed-needy, and
1049 * unlikely to need full 64-bitness. */
1050 U8 digit = *s++ - '0';
945b524a
JH
1051 if (val < uv_max_div_10 ||
1052 (val == uv_max_div_10 && digit <= uv_max_mod_10)) {
75feedba
JH
1053 val = val * 10 + digit;
1054 } else {
6313e544 1055 *eptr = NULL;
75feedba 1056 return UV_MAX;
6313e544 1057 }
6313e544
JH
1058 }
1059 }
75feedba
JH
1060 }
1061 if (s == pv) {
1062 *eptr = NULL; /* If no progress, failed to parse anything. */
1063 return 0;
6313e544
JH
1064 }
1065 if (endptr == NULL && *s) {
1066 return 0; /* If endptr is NULL, no trailing non-digits allowed. */
1067 }
1068 *eptr = s;
1069 return val;
1070}
1071
a4eca1d4 1072#ifndef USE_QUADMATH
4801ca72 1073STATIC NV
98994639
HS
1074S_mulexp10(NV value, I32 exponent)
1075{
1076 NV result = 1.0;
1077 NV power = 10.0;
1078 bool negative = 0;
1079 I32 bit;
1080
1081 if (exponent == 0)
1082 return value;
659c4b96
DM
1083 if (value == 0)
1084 return (NV)0;
87032ba1 1085
24866caa 1086 /* On OpenVMS VAX we by default use the D_FLOAT double format,
67597c89 1087 * and that format does not have *easy* capabilities [1] for
24866caa
CB
1088 * overflowing doubles 'silently' as IEEE fp does. We also need
1089 * to support G_FLOAT on both VAX and Alpha, and though the exponent
1090 * range is much larger than D_FLOAT it still doesn't do silent
1091 * overflow. Therefore we need to detect early whether we would
1092 * overflow (this is the behaviour of the native string-to-float
1093 * conversion routines, and therefore of native applications, too).
67597c89 1094 *
24866caa
CB
1095 * [1] Trying to establish a condition handler to trap floating point
1096 * exceptions is not a good idea. */
87032ba1
JH
1097
1098 /* In UNICOS and in certain Cray models (such as T90) there is no
1099 * IEEE fp, and no way at all from C to catch fp overflows gracefully.
1100 * There is something you can do if you are willing to use some
1101 * inline assembler: the instruction is called DFI-- but that will
1102 * disable *all* floating point interrupts, a little bit too large
1103 * a hammer. Therefore we need to catch potential overflows before
1104 * it's too late. */
353813d9 1105
85bba25f 1106#if ((defined(VMS) && !defined(_IEEE_FP)) || defined(_UNICOS)) && defined(NV_MAX_10_EXP)
353813d9 1107 STMT_START {
c4420975 1108 const NV exp_v = log10(value);
353813d9
HS
1109 if (exponent >= NV_MAX_10_EXP || exponent + exp_v >= NV_MAX_10_EXP)
1110 return NV_MAX;
1111 if (exponent < 0) {
1112 if (-(exponent + exp_v) >= NV_MAX_10_EXP)
1113 return 0.0;
1114 while (-exponent >= NV_MAX_10_EXP) {
1115 /* combination does not overflow, but 10^(-exponent) does */
1116 value /= 10;
1117 ++exponent;
1118 }
1119 }
1120 } STMT_END;
87032ba1
JH
1121#endif
1122
353813d9
HS
1123 if (exponent < 0) {
1124 negative = 1;
1125 exponent = -exponent;
b27804d8
DM
1126#ifdef NV_MAX_10_EXP
1127 /* for something like 1234 x 10^-309, the action of calculating
1128 * the intermediate value 10^309 then returning 1234 / (10^309)
1129 * will fail, since 10^309 becomes infinity. In this case try to
1130 * refactor it as 123 / (10^308) etc.
1131 */
1132 while (value && exponent > NV_MAX_10_EXP) {
1133 exponent--;
1134 value /= 10;
1135 }
48853916
JH
1136 if (value == 0.0)
1137 return value;
b27804d8 1138#endif
353813d9 1139 }
c62e754c
JH
1140#if defined(__osf__)
1141 /* Even with cc -ieee + ieee_set_fp_control(IEEE_TRAP_ENABLE_INV)
1142 * Tru64 fp behavior on inf/nan is somewhat broken. Another way
1143 * to do this would be ieee_set_fp_control(IEEE_TRAP_ENABLE_OVF)
1144 * but that breaks another set of infnan.t tests. */
1145# define FP_OVERFLOWS_TO_ZERO
1146#endif
98994639
HS
1147 for (bit = 1; exponent; bit <<= 1) {
1148 if (exponent & bit) {
1149 exponent ^= bit;
1150 result *= power;
c62e754c
JH
1151#ifdef FP_OVERFLOWS_TO_ZERO
1152 if (result == 0)
1153 return value < 0 ? -NV_INF : NV_INF;
1154#endif
236f0012
CB
1155 /* Floating point exceptions are supposed to be turned off,
1156 * but if we're obviously done, don't risk another iteration.
1157 */
1158 if (exponent == 0) break;
98994639
HS
1159 }
1160 power *= power;
1161 }
1162 return negative ? value / result : value * result;
1163}
a4eca1d4 1164#endif /* #ifndef USE_QUADMATH */
98994639
HS
1165
1166NV
1167Perl_my_atof(pTHX_ const char* s)
1168{
1169 NV x = 0.0;
a4eca1d4
JH
1170#ifdef USE_QUADMATH
1171 Perl_my_atof2(aTHX_ s, &x);
1172 return x;
1173#else
1174# ifdef USE_LOCALE_NUMERIC
7918f24d
NC
1175 PERL_ARGS_ASSERT_MY_ATOF;
1176
a2287a13
KW
1177 {
1178 DECLARE_STORE_LC_NUMERIC_SET_TO_NEEDED();
d6ded950 1179 if (PL_numeric_radix_sv && IN_LC(LC_NUMERIC)) {
e4850248
KW
1180 const char *standard = NULL, *local = NULL;
1181 bool use_standard_radix;
98994639 1182
e4850248
KW
1183 /* Look through the string for the first thing that looks like a
1184 * decimal point: either the value in the current locale or the
1185 * standard fallback of '.'. The one which appears earliest in the
1186 * input string is the one that we should have atof look for. Note
1187 * that we have to determine this beforehand because on some
1188 * systems, Perl_atof2 is just a wrapper around the system's atof.
1189 * */
1190 standard = strchr(s, '.');
1191 local = strstr(s, SvPV_nolen(PL_numeric_radix_sv));
78787052 1192
e4850248 1193 use_standard_radix = standard && (!local || standard < local);
78787052 1194
e4850248
KW
1195 if (use_standard_radix)
1196 SET_NUMERIC_STANDARD();
78787052 1197
e4850248 1198 Perl_atof2(s, x);
78787052 1199
e4850248
KW
1200 if (use_standard_radix)
1201 SET_NUMERIC_LOCAL();
1202 }
1203 else
1204 Perl_atof2(s, x);
a2287a13
KW
1205 RESTORE_LC_NUMERIC();
1206 }
a4eca1d4 1207# else
a36244b7 1208 Perl_atof2(s, x);
a4eca1d4 1209# endif
98994639
HS
1210#endif
1211 return x;
1212}
1213
829757a4
JH
1214static char*
1215S_my_atof_infnan(const char* s, bool negative, const char* send, NV* value)
1216{
1217 const char *p0 = negative ? s - 1 : s;
1218 const char *p = p0;
1219 int infnan = grok_infnan(&p, send);
1220 if (infnan && p != p0) {
1221 /* If we can generate inf/nan directly, let's do so. */
1222#ifdef NV_INF
1223 if ((infnan & IS_NUMBER_INFINITY)) {
1224 *value = (infnan & IS_NUMBER_NEG) ? -NV_INF: NV_INF;
1225 return (char*)p;
1226 }
1227#endif
1228#ifdef NV_NAN
1229 if ((infnan & IS_NUMBER_NAN)) {
1230 *value = NV_NAN;
1231 return (char*)p;
1232 }
1233#endif
1234#ifdef Perl_strtod
68611e6f 1235 /* If still here, we didn't have either NV_INF or NV_NAN,
829757a4
JH
1236 * and can try falling back to native strtod/strtold.
1237 *
68611e6f
JH
1238 * (Though, are our NV_INF or NV_NAN ever not defined?)
1239 *
829757a4
JH
1240 * The native interface might not recognize all the possible
1241 * inf/nan strings Perl recognizes. What we can try
1242 * is to try faking the input. We will try inf/-inf/nan
1243 * as the most promising/portable input. */
1244 {
1245 const char* fake = NULL;
1246 char* endp;
1247 NV nv;
1248 if ((infnan & IS_NUMBER_INFINITY)) {
1249 fake = ((infnan & IS_NUMBER_NEG)) ? "-inf" : "inf";
1250 }
1251 else if ((infnan & IS_NUMBER_NAN)) {
1252 fake = "nan";
1253 }
1254 assert(fake);
1255 nv = Perl_strtod(fake, &endp);
1256 if (fake != endp) {
1257 if ((infnan & IS_NUMBER_INFINITY)) {
1258#ifdef Perl_isinf
1259 if (Perl_isinf(nv))
1260 *value = nv;
1261#else
1262 /* last resort, may generate SIGFPE */
1263 *value = Perl_exp((NV)1e9);
1264 if ((infnan & IS_NUMBER_NEG))
1265 *value = -*value;
1266#endif
1267 return (char*)p; /* p, not endp */
1268 }
1269 else if ((infnan & IS_NUMBER_NAN)) {
1270#ifdef Perl_isnan
1271 if (Perl_isnan(nv))
1272 *value = nv;
1273#else
1274 /* last resort, may generate SIGFPE */
1275 *value = Perl_log((NV)-1.0);
1276#endif
1277 return (char*)p; /* p, not endp */
1278 }
1279 }
1280 }
1281#endif /* #ifdef Perl_strtod */
1282 }
1283 return NULL;
1284}
1285
98994639
HS
1286char*
1287Perl_my_atof2(pTHX_ const char* orig, NV* value)
1288{
e1ec3a88 1289 const char* s = orig;
a4eca1d4
JH
1290 NV result[3] = {0.0, 0.0, 0.0};
1291#if defined(USE_PERL_ATOF) || defined(USE_QUADMATH)
ae776a2c 1292 const char* send = s + strlen(orig); /* one past the last */
a4eca1d4
JH
1293 bool negative = 0;
1294#endif
1295#if defined(USE_PERL_ATOF) && !defined(USE_QUADMATH)
1296 UV accumulator[2] = {0,0}; /* before/after dp */
8194bf88 1297 bool seen_digit = 0;
20f6aaab
AS
1298 I32 exp_adjust[2] = {0,0};
1299 I32 exp_acc[2] = {-1, -1};
1300 /* the current exponent adjust for the accumulators */
98994639 1301 I32 exponent = 0;
8194bf88 1302 I32 seen_dp = 0;
20f6aaab
AS
1303 I32 digit = 0;
1304 I32 old_digit = 0;
8194bf88 1305 I32 sig_digits = 0; /* noof significant digits seen so far */
a4eca1d4 1306#endif
8194bf88 1307
a4eca1d4 1308#if defined(USE_PERL_ATOF) || defined(USE_QUADMATH)
7918f24d
NC
1309 PERL_ARGS_ASSERT_MY_ATOF2;
1310
a4eca1d4
JH
1311 /* leading whitespace */
1312 while (isSPACE(*s))
1313 ++s;
1314
1315 /* sign */
1316 switch (*s) {
1317 case '-':
1318 negative = 1;
1319 /* FALLTHROUGH */
1320 case '+':
1321 ++s;
1322 }
1323#endif
1324
1325#ifdef USE_QUADMATH
1326 {
1327 char* endp;
1328 if ((endp = S_my_atof_infnan(s, negative, send, value)))
1329 return endp;
1330 result[2] = strtoflt128(s, &endp);
1331 if (s != endp) {
1332 *value = negative ? -result[2] : result[2];
1333 return endp;
1334 }
1335 return NULL;
1336 }
1337#elif defined(USE_PERL_ATOF)
1338
8194bf88
DM
1339/* There is no point in processing more significant digits
1340 * than the NV can hold. Note that NV_DIG is a lower-bound value,
1341 * while we need an upper-bound value. We add 2 to account for this;
1342 * since it will have been conservative on both the first and last digit.
1343 * For example a 32-bit mantissa with an exponent of 4 would have
1344 * exact values in the set
1345 * 4
1346 * 8
1347 * ..
1348 * 17179869172
1349 * 17179869176
1350 * 17179869180
1351 *
1352 * where for the purposes of calculating NV_DIG we would have to discount
1353 * both the first and last digit, since neither can hold all values from
1354 * 0..9; but for calculating the value we must examine those two digits.
1355 */
ffa277e5
EAS
1356#ifdef MAX_SIG_DIG_PLUS
1357 /* It is not necessarily the case that adding 2 to NV_DIG gets all the
1358 possible digits in a NV, especially if NVs are not IEEE compliant
1359 (e.g., long doubles on IRIX) - Allen <allens@cpan.org> */
1360# define MAX_SIG_DIGITS (NV_DIG+MAX_SIG_DIG_PLUS)
1361#else
1362# define MAX_SIG_DIGITS (NV_DIG+2)
1363#endif
8194bf88
DM
1364
1365/* the max number we can accumulate in a UV, and still safely do 10*N+9 */
1366#define MAX_ACCUMULATE ( (UV) ((UV_MAX - 9)/10))
98994639 1367
ae776a2c 1368 {
829757a4
JH
1369 const char* endp;
1370 if ((endp = S_my_atof_infnan(s, negative, send, value)))
1371 return (char*)endp;
ae776a2c 1372 }
2b54f59f 1373
8194bf88
DM
1374 /* we accumulate digits into an integer; when this becomes too
1375 * large, we add the total to NV and start again */
98994639 1376
8194bf88
DM
1377 while (1) {
1378 if (isDIGIT(*s)) {
1379 seen_digit = 1;
20f6aaab 1380 old_digit = digit;
8194bf88 1381 digit = *s++ - '0';
20f6aaab
AS
1382 if (seen_dp)
1383 exp_adjust[1]++;
98994639 1384
8194bf88
DM
1385 /* don't start counting until we see the first significant
1386 * digit, eg the 5 in 0.00005... */
1387 if (!sig_digits && digit == 0)
1388 continue;
1389
1390 if (++sig_digits > MAX_SIG_DIGITS) {
98994639 1391 /* limits of precision reached */
20f6aaab
AS
1392 if (digit > 5) {
1393 ++accumulator[seen_dp];
1394 } else if (digit == 5) {
1395 if (old_digit % 2) { /* round to even - Allen */
1396 ++accumulator[seen_dp];
1397 }
1398 }
1399 if (seen_dp) {
1400 exp_adjust[1]--;
1401 } else {
1402 exp_adjust[0]++;
1403 }
8194bf88 1404 /* skip remaining digits */
98994639 1405 while (isDIGIT(*s)) {
98994639 1406 ++s;
20f6aaab
AS
1407 if (! seen_dp) {
1408 exp_adjust[0]++;
1409 }
98994639
HS
1410 }
1411 /* warn of loss of precision? */
98994639 1412 }
8194bf88 1413 else {
20f6aaab 1414 if (accumulator[seen_dp] > MAX_ACCUMULATE) {
8194bf88 1415 /* add accumulator to result and start again */
20f6aaab
AS
1416 result[seen_dp] = S_mulexp10(result[seen_dp],
1417 exp_acc[seen_dp])
1418 + (NV)accumulator[seen_dp];
1419 accumulator[seen_dp] = 0;
1420 exp_acc[seen_dp] = 0;
98994639 1421 }
20f6aaab
AS
1422 accumulator[seen_dp] = accumulator[seen_dp] * 10 + digit;
1423 ++exp_acc[seen_dp];
98994639 1424 }
8194bf88 1425 }
e1ec3a88 1426 else if (!seen_dp && GROK_NUMERIC_RADIX(&s, send)) {
8194bf88 1427 seen_dp = 1;
20f6aaab 1428 if (sig_digits > MAX_SIG_DIGITS) {
c86f7df5 1429 do {
20f6aaab 1430 ++s;
c86f7df5 1431 } while (isDIGIT(*s));
20f6aaab
AS
1432 break;
1433 }
8194bf88
DM
1434 }
1435 else {
1436 break;
98994639
HS
1437 }
1438 }
1439
20f6aaab
AS
1440 result[0] = S_mulexp10(result[0], exp_acc[0]) + (NV)accumulator[0];
1441 if (seen_dp) {
1442 result[1] = S_mulexp10(result[1], exp_acc[1]) + (NV)accumulator[1];
1443 }
98994639 1444
305b8651 1445 if (seen_digit && (isALPHA_FOLD_EQ(*s, 'e'))) {
98994639
HS
1446 bool expnegative = 0;
1447
1448 ++s;
1449 switch (*s) {
1450 case '-':
1451 expnegative = 1;
924ba076 1452 /* FALLTHROUGH */
98994639
HS
1453 case '+':
1454 ++s;
1455 }
1456 while (isDIGIT(*s))
1457 exponent = exponent * 10 + (*s++ - '0');
1458 if (expnegative)
1459 exponent = -exponent;
1460 }
1461
20f6aaab
AS
1462
1463
98994639 1464 /* now apply the exponent */
20f6aaab
AS
1465
1466 if (seen_dp) {
1467 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0])
1468 + S_mulexp10(result[1],exponent-exp_adjust[1]);
1469 } else {
1470 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0]);
1471 }
98994639
HS
1472
1473 /* now apply the sign */
1474 if (negative)
20f6aaab 1475 result[2] = -result[2];
a36244b7 1476#endif /* USE_PERL_ATOF */
20f6aaab 1477 *value = result[2];
73d840c0 1478 return (char *)s;
98994639
HS
1479}
1480
5d34af89 1481/*
3d9d9213 1482=for apidoc isinfnan
5d34af89
JH
1483
1484Perl_isinfnan() is utility function that returns true if the NV
1485argument is either an infinity or a NaN, false otherwise. To test
1486in more detail, use Perl_isinf() and Perl_isnan().
1487
68611e6f
JH
1488This is also the logical inverse of Perl_isfinite().
1489
5d34af89
JH
1490=cut
1491*/
1cd88304
JH
1492bool
1493Perl_isinfnan(NV nv)
1494{
1495#ifdef Perl_isinf
1496 if (Perl_isinf(nv))
1497 return TRUE;
1498#endif
1499#ifdef Perl_isnan
1500 if (Perl_isnan(nv))
1501 return TRUE;
1502#endif
1503 return FALSE;
1504}
1505
354b74ae
FC
1506/*
1507=for apidoc
1508
1509Checks whether the argument would be either an infinity or NaN when used
1510as a number, but is careful not to trigger non-numeric or uninitialized
1511warnings. it assumes the caller has done SvGETMAGIC(sv) already.
1512
1513=cut
1514*/
1515
1516bool
1517Perl_isinfnansv(pTHX_ SV *sv)
1518{
1519 PERL_ARGS_ASSERT_ISINFNANSV;
1520 if (!SvOK(sv))
1521 return FALSE;
1522 if (SvNOKp(sv))
1523 return Perl_isinfnan(SvNVX(sv));
1524 if (SvIOKp(sv))
1525 return FALSE;
1526 {
1527 STRLEN len;
1528 const char *s = SvPV_nomg_const(sv, len);
1529 return cBOOL(grok_infnan(&s, s+len));
1530 }
1531}
1532
d67dac15 1533#ifndef HAS_MODFL
68611e6f
JH
1534/* C99 has truncl, pre-C99 Solaris had aintl. We can use either with
1535 * copysignl to emulate modfl, which is in some platforms missing or
1536 * broken. */
d67dac15
JH
1537# if defined(HAS_TRUNCL) && defined(HAS_COPYSIGNL)
1538long double
1539Perl_my_modfl(long double x, long double *ip)
1540{
68611e6f
JH
1541 *ip = truncl(x);
1542 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
d67dac15
JH
1543}
1544# elif defined(HAS_AINTL) && defined(HAS_COPYSIGNL)
55954f19
JH
1545long double
1546Perl_my_modfl(long double x, long double *ip)
1547{
68611e6f
JH
1548 *ip = aintl(x);
1549 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
55954f19 1550}
d67dac15 1551# endif
55954f19
JH
1552#endif
1553
7b9b7dff 1554/* Similarly, with ilogbl and scalbnl we can emulate frexpl. */
55954f19
JH
1555#if ! defined(HAS_FREXPL) && defined(HAS_ILOGBL) && defined(HAS_SCALBNL)
1556long double
1557Perl_my_frexpl(long double x, int *e) {
68611e6f
JH
1558 *e = x == 0.0L ? 0 : ilogbl(x) + 1;
1559 return (scalbnl(x, -*e));
55954f19
JH
1560}
1561#endif
66610fdd
RGS
1562
1563/*
ed140128
AD
1564=for apidoc Perl_signbit
1565
1566Return a non-zero integer if the sign bit on an NV is set, and 0 if
1567it is not.
1568
1569If Configure detects this system has a signbit() that will work with
1570our NVs, then we just use it via the #define in perl.h. Otherwise,
8b7fad81
JH
1571fall back on this implementation. The main use of this function
1572is catching -0.0.
ed140128
AD
1573
1574Configure notes: This function is called 'Perl_signbit' instead of a
1575plain 'signbit' because it is easy to imagine a system having a signbit()
1576function or macro that doesn't happen to work with our particular choice
1577of NVs. We shouldn't just re-#define signbit as Perl_signbit and expect
1578the standard system headers to be happy. Also, this is a no-context
1579function (no pTHX_) because Perl_signbit() is usually re-#defined in
1580perl.h as a simple macro call to the system's signbit().
1581Users should just always call Perl_signbit().
1582
1583=cut
1584*/
1585#if !defined(HAS_SIGNBIT)
1586int
1587Perl_signbit(NV x) {
8b7fad81
JH
1588# ifdef Perl_fp_class_nzero
1589 if (x == 0)
1590 return Perl_fp_class_nzero(x);
8b7fad81 1591# endif
3585840c 1592 return (x < 0.0) ? 1 : 0;
ed140128
AD
1593}
1594#endif
1595
1596/*
66610fdd
RGS
1597 * Local variables:
1598 * c-indentation-style: bsd
1599 * c-basic-offset: 4
14d04a33 1600 * indent-tabs-mode: nil
66610fdd
RGS
1601 * End:
1602 *
14d04a33 1603 * ex: set ts=8 sts=4 sw=4 et:
37442d52 1604 */