This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
[perl #123814] stricter handling of numbers in regexp quantifiers
[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
569f27e5 550/*
ff4eb398
JH
551=for apidoc grok_infnan
552
553Helper for grok_number(), accepts various ways of spelling "infinity"
554or "not a number", and returns one of the following flag combinations:
555
556 IS_NUMBER_INFINITE
557 IS_NUMBER_NAN
558 IS_NUMBER_INFINITE | IS_NUMBER_NEG
559 IS_NUMBER_NAN | IS_NUMBER_NEG
560 0
561
62bdc035 562possibly |-ed with IS_NUMBER_TRAILING.
b489e20f 563
62bdc035
JH
564If an infinity or a not-a-number is recognized, the *sp will point to
565one byte past the end of the recognized string. If the recognition fails,
ff4eb398
JH
566zero is returned, and the *sp will not move.
567
568=cut
569*/
570
571int
3823048b 572Perl_grok_infnan(pTHX_ const char** sp, const char* send)
ff4eb398
JH
573{
574 const char* s = *sp;
575 int flags = 0;
62bdc035 576 bool odh = FALSE; /* one-dot-hash: 1.#INF */
ff4eb398
JH
577
578 PERL_ARGS_ASSERT_GROK_INFNAN;
579
8c12dc63
JH
580 if (*s == '+') {
581 s++; if (s == send) return 0;
582 }
583 else if (*s == '-') {
ff4eb398
JH
584 flags |= IS_NUMBER_NEG; /* Yes, -NaN happens. Incorrect but happens. */
585 s++; if (s == send) return 0;
586 }
587
588 if (*s == '1') {
62bdc035
JH
589 /* Visual C: 1.#SNAN, -1.#QNAN, 1#INF, 1.#IND (maybe also 1.#NAN)
590 * Let's keep the dot optional. */
ff4eb398
JH
591 s++; if (s == send) return 0;
592 if (*s == '.') {
593 s++; if (s == send) return 0;
594 }
595 if (*s == '#') {
596 s++; if (s == send) return 0;
597 } else
598 return 0;
e855f543 599 odh = TRUE;
ff4eb398
JH
600 }
601
305b8651 602 if (isALPHA_FOLD_EQ(*s, 'I')) {
62bdc035
JH
603 /* INF or IND (1.#IND is "indeterminate", a certain type of NAN) */
604
305b8651 605 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
ff4eb398 606 s++; if (s == send) return 0;
305b8651 607 if (isALPHA_FOLD_EQ(*s, 'F')) {
ff4eb398 608 s++;
b8974fcb
JH
609 if (s < send && (isALPHA_FOLD_EQ(*s, 'I'))) {
610 int fail =
611 flags | IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT | IS_NUMBER_TRAILING;
612 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return fail;
613 s++; if (s == send || isALPHA_FOLD_NE(*s, 'I')) return fail;
614 s++; if (s == send || isALPHA_FOLD_NE(*s, 'T')) return fail;
615 s++; if (s == send || isALPHA_FOLD_NE(*s, 'Y')) return fail;
3396ed30 616 s++;
b8974fcb
JH
617 } else if (odh) {
618 while (*s == '0') { /* 1.#INF00 */
619 s++;
620 }
3396ed30 621 }
b489e20f
JH
622 while (s < send && isSPACE(*s))
623 s++;
624 if (s < send && *s) {
3396ed30 625 flags |= IS_NUMBER_TRAILING;
fae4db12 626 }
ff4eb398
JH
627 flags |= IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT;
628 }
e855f543 629 else if (isALPHA_FOLD_EQ(*s, 'D') && odh) { /* 1.#IND */
ff4eb398
JH
630 s++;
631 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
fae4db12
JH
632 while (*s == '0') { /* 1.#IND00 */
633 s++;
634 }
1e9aa12f
JH
635 if (*s) {
636 flags |= IS_NUMBER_TRAILING;
637 }
ff4eb398
JH
638 } else
639 return 0;
ff4eb398
JH
640 }
641 else {
62bdc035 642 /* Maybe NAN of some sort */
3823048b
JH
643
644 if (isALPHA_FOLD_EQ(*s, 'S') || isALPHA_FOLD_EQ(*s, 'Q')) {
645 /* snan, qNaN */
646 /* XXX do something with the snan/qnan difference */
647 s++; if (s == send) return 0;
648 }
649
650 if (isALPHA_FOLD_EQ(*s, 'N')) {
651 s++; if (s == send || isALPHA_FOLD_NE(*s, 'A')) return 0;
652 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
653 s++;
654
655 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
656
657 /* NaN can be followed by various stuff (NaNQ, NaNS), but
658 * there are also multiple different NaN values, and some
659 * implementations output the "payload" values,
660 * e.g. NaN123, NAN(abc), while some legacy implementations
661 * have weird stuff like NaN%. */
662 if (isALPHA_FOLD_EQ(*s, 'q') ||
663 isALPHA_FOLD_EQ(*s, 's')) {
664 /* "nanq" or "nans" are ok, though generating
665 * these portably is tricky. */
666 s++;
667 }
668 if (*s == '(') {
669 /* C99 style "nan(123)" or Perlish equivalent "nan($uv)". */
670 const char *t;
671 s++;
672 if (s == send) {
673 return flags | IS_NUMBER_TRAILING;
674 }
675 t = s + 1;
676 while (t < send && *t && *t != ')') {
677 t++;
678 }
679 if (t == send) {
680 return flags | IS_NUMBER_TRAILING;
681 }
682 if (*t == ')') {
683 int nantype;
684 UV nanval;
685 if (s[0] == '0' && s + 2 < t &&
686 isALPHA_FOLD_EQ(s[1], 'x') &&
687 isXDIGIT(s[2])) {
688 STRLEN len = t - s;
689 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES;
690 nanval = grok_hex(s, &len, &flags, NULL);
691 if ((flags & PERL_SCAN_GREATER_THAN_UV_MAX)) {
692 nantype = 0;
693 } else {
694 nantype = IS_NUMBER_IN_UV;
695 }
696 s += len;
697 } else if (s[0] == '0' && s + 2 < t &&
698 isALPHA_FOLD_EQ(s[1], 'b') &&
699 (s[2] == '0' || s[2] == '1')) {
700 STRLEN len = t - s;
701 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES;
702 nanval = grok_bin(s, &len, &flags, NULL);
703 if ((flags & PERL_SCAN_GREATER_THAN_UV_MAX)) {
704 nantype = 0;
705 } else {
706 nantype = IS_NUMBER_IN_UV;
707 }
708 s += len;
709 } else {
710 const char *u;
711 nantype =
712 grok_number_flags(s, t - s, &nanval,
713 PERL_SCAN_TRAILING |
714 PERL_SCAN_ALLOW_UNDERSCORES);
715 /* Unfortunately grok_number_flags() doesn't
716 * tell how far we got and the ')' will always
717 * be "trailing", so we need to double-check
718 * whether we had something dubious. */
719 for (u = s; u < t; u++) {
720 if (!isDIGIT(*u)) {
721 flags |= IS_NUMBER_TRAILING;
722 break;
723 }
724 }
725 s = u;
726 }
727
728 /* XXX Doesn't do octal: nan("0123").
729 * Probably not a big loss. */
730
731 if ((nantype & IS_NUMBER_NOT_INT) ||
732 !(nantype && IS_NUMBER_IN_UV)) {
733 /* XXX the nanval is currently unused, that is,
734 * not inserted as the NaN payload of the NV.
735 * But the above code already parses the C99
736 * nan(...) format. See below, and see also
737 * the nan() in POSIX.xs.
738 *
739 * Certain configuration combinations where
740 * NVSIZE is greater than UVSIZE mean that
741 * a single UV cannot contain all the possible
742 * NaN payload bits. There would need to be
743 * some more generic syntax than "nan($uv)".
744 *
745 * Issues to keep in mind:
746 *
747 * (1) In most common cases there would
748 * not be an integral number of bytes that
749 * could be set, only a certain number of bits.
750 * For example for the common case of
751 * NVSIZE == UVSIZE == 8 there is room for 52
752 * bits in the payload, but the most significant
753 * bit is commonly reserved for the
754 * signaling/quiet bit, leaving 51 bits.
755 * Furthermore, the C99 nan() is supposed
756 * to generate quiet NaNs, so it is doubtful
757 * whether it should be able to generate
758 * signaling NaNs. For the x86 80-bit doubles
759 * (if building a long double Perl) there would
760 * be 62 bits (s/q bit being the 63rd).
761 *
762 * (2) Endianness of the payload bits. If the
763 * payload is specified as an UV, the low-order
764 * bits of the UV are naturally little-endianed
765 * (rightmost) bits of the payload. The endianness
766 * of UVs and NVs can be different. */
767 return 0;
768 }
769 if (s < t) {
770 flags |= IS_NUMBER_TRAILING;
771 }
772 } else {
773 /* Looked like nan(...), but no close paren. */
774 flags |= IS_NUMBER_TRAILING;
775 }
776 } else {
777 while (s < send && isSPACE(*s))
778 s++;
779 if (s < send && *s) {
780 /* Note that we here implicitly accept (parse as
781 * "nan", but with warnings) also any other weird
782 * trailing stuff for "nan". In the above we just
783 * check that if we got the C99-style "nan(...)",
784 * the "..." looks sane.
785 * If in future we accept more ways of specifying
786 * the nan payload, the accepting would happen around
787 * here. */
788 flags |= IS_NUMBER_TRAILING;
789 }
790 }
791 s = send;
792 }
793 else
794 return 0;
ff4eb398
JH
795 }
796
b489e20f
JH
797 while (s < send && isSPACE(*s))
798 s++;
799
a1fe7cea
JH
800 *sp = s;
801 return flags;
ff4eb398
JH
802}
803
13393a5e 804/*
3823048b 805=for apidoc grok_number_flags
13393a5e
JH
806
807Recognise (or not) a number. The type of the number is returned
808(0 if unrecognised), otherwise it is a bit-ORed combination of
809IS_NUMBER_IN_UV, IS_NUMBER_GREATER_THAN_UV_MAX, IS_NUMBER_NOT_INT,
810IS_NUMBER_NEG, IS_NUMBER_INFINITY, IS_NUMBER_NAN (defined in perl.h).
811
812If the value of the number can fit in a UV, it is returned in the *valuep
813IS_NUMBER_IN_UV will be set to indicate that *valuep is valid, IS_NUMBER_IN_UV
814will never be set unless *valuep is valid, but *valuep may have been assigned
815to during processing even though IS_NUMBER_IN_UV is not set on return.
816If valuep is NULL, IS_NUMBER_IN_UV will be set for the same cases as when
817valuep is non-NULL, but no actual assignment (or SEGV) will occur.
818
819IS_NUMBER_NOT_INT will be set with IS_NUMBER_IN_UV if trailing decimals were
820seen (in which case *valuep gives the true value truncated to an integer), and
821IS_NUMBER_NEG if the number is negative (in which case *valuep holds the
822absolute value). IS_NUMBER_IN_UV is not set if e notation was used or the
823number is larger than a UV.
824
825C<flags> allows only C<PERL_SCAN_TRAILING>, which allows for trailing
826non-numeric text on an otherwise successful I<grok>, setting
827C<IS_NUMBER_TRAILING> on the result.
828
829=for apidoc grok_number
830
831Identical to grok_number_flags() with flags set to zero.
832
833=cut
834 */
835int
836Perl_grok_number(pTHX_ const char *pv, STRLEN len, UV *valuep)
837{
838 PERL_ARGS_ASSERT_GROK_NUMBER;
839
840 return grok_number_flags(pv, len, valuep, 0);
841}
842
945b524a
JH
843static const UV uv_max_div_10 = UV_MAX / 10;
844static const U8 uv_max_mod_10 = UV_MAX % 10;
845
3f7602fa 846int
3823048b 847Perl_grok_number_flags(pTHX_ const char *pv, STRLEN len, UV *valuep, U32 flags)
3f7602fa 848{
60939fb8 849 const char *s = pv;
c4420975 850 const char * const send = pv + len;
ae776a2c 851 const char *d;
60939fb8 852 int numtype = 0;
60939fb8 853
3823048b 854 PERL_ARGS_ASSERT_GROK_NUMBER_FLAGS;
7918f24d 855
60939fb8
NC
856 while (s < send && isSPACE(*s))
857 s++;
858 if (s == send) {
859 return 0;
860 } else if (*s == '-') {
861 s++;
862 numtype = IS_NUMBER_NEG;
863 }
864 else if (*s == '+')
aa42a541 865 s++;
60939fb8
NC
866
867 if (s == send)
868 return 0;
869
ae776a2c 870 /* The first digit (after optional sign): note that might
8c12dc63 871 * also point to "infinity" or "nan", or "1.#INF". */
ae776a2c
JH
872 d = s;
873
8c12dc63 874 /* next must be digit or the radix separator or beginning of infinity/nan */
60939fb8
NC
875 if (isDIGIT(*s)) {
876 /* UVs are at least 32 bits, so the first 9 decimal digits cannot
877 overflow. */
878 UV value = *s - '0';
879 /* This construction seems to be more optimiser friendly.
880 (without it gcc does the isDIGIT test and the *s - '0' separately)
881 With it gcc on arm is managing 6 instructions (6 cycles) per digit.
882 In theory the optimiser could deduce how far to unroll the loop
883 before checking for overflow. */
58bb9ec3
NC
884 if (++s < send) {
885 int digit = *s - '0';
60939fb8
NC
886 if (digit >= 0 && digit <= 9) {
887 value = value * 10 + digit;
58bb9ec3
NC
888 if (++s < send) {
889 digit = *s - '0';
60939fb8
NC
890 if (digit >= 0 && digit <= 9) {
891 value = value * 10 + digit;
58bb9ec3
NC
892 if (++s < send) {
893 digit = *s - '0';
60939fb8
NC
894 if (digit >= 0 && digit <= 9) {
895 value = value * 10 + digit;
58bb9ec3
NC
896 if (++s < send) {
897 digit = *s - '0';
60939fb8
NC
898 if (digit >= 0 && digit <= 9) {
899 value = value * 10 + digit;
58bb9ec3
NC
900 if (++s < send) {
901 digit = *s - '0';
60939fb8
NC
902 if (digit >= 0 && digit <= 9) {
903 value = value * 10 + digit;
58bb9ec3
NC
904 if (++s < send) {
905 digit = *s - '0';
60939fb8
NC
906 if (digit >= 0 && digit <= 9) {
907 value = value * 10 + digit;
58bb9ec3
NC
908 if (++s < send) {
909 digit = *s - '0';
60939fb8
NC
910 if (digit >= 0 && digit <= 9) {
911 value = value * 10 + digit;
58bb9ec3
NC
912 if (++s < send) {
913 digit = *s - '0';
60939fb8
NC
914 if (digit >= 0 && digit <= 9) {
915 value = value * 10 + digit;
58bb9ec3 916 if (++s < send) {
60939fb8
NC
917 /* Now got 9 digits, so need to check
918 each time for overflow. */
58bb9ec3 919 digit = *s - '0';
60939fb8 920 while (digit >= 0 && digit <= 9
945b524a
JH
921 && (value < uv_max_div_10
922 || (value == uv_max_div_10
923 && digit <= uv_max_mod_10))) {
60939fb8 924 value = value * 10 + digit;
58bb9ec3
NC
925 if (++s < send)
926 digit = *s - '0';
60939fb8
NC
927 else
928 break;
929 }
930 if (digit >= 0 && digit <= 9
51bd16da 931 && (s < send)) {
60939fb8
NC
932 /* value overflowed.
933 skip the remaining digits, don't
934 worry about setting *valuep. */
935 do {
936 s++;
937 } while (s < send && isDIGIT(*s));
938 numtype |=
939 IS_NUMBER_GREATER_THAN_UV_MAX;
940 goto skip_value;
941 }
942 }
943 }
98994639 944 }
60939fb8
NC
945 }
946 }
947 }
948 }
949 }
950 }
951 }
952 }
953 }
954 }
955 }
98994639 956 }
60939fb8 957 }
98994639 958 }
60939fb8
NC
959 numtype |= IS_NUMBER_IN_UV;
960 if (valuep)
961 *valuep = value;
962
963 skip_value:
964 if (GROK_NUMERIC_RADIX(&s, send)) {
965 numtype |= IS_NUMBER_NOT_INT;
966 while (s < send && isDIGIT(*s)) /* optional digits after the radix */
967 s++;
98994639 968 }
60939fb8
NC
969 }
970 else if (GROK_NUMERIC_RADIX(&s, send)) {
971 numtype |= IS_NUMBER_NOT_INT | IS_NUMBER_IN_UV; /* valuep assigned below */
972 /* no digits before the radix means we need digits after it */
973 if (s < send && isDIGIT(*s)) {
974 do {
975 s++;
976 } while (s < send && isDIGIT(*s));
977 if (valuep) {
978 /* integer approximation is valid - it's 0. */
979 *valuep = 0;
980 }
98994639 981 }
60939fb8 982 else
ae776a2c 983 return 0;
ff4eb398 984 }
60939fb8 985
926f5fc6 986 if (s > d && s < send) {
60939fb8 987 /* we can have an optional exponent part */
305b8651 988 if (isALPHA_FOLD_EQ(*s, 'e')) {
60939fb8
NC
989 s++;
990 if (s < send && (*s == '-' || *s == '+'))
991 s++;
992 if (s < send && isDIGIT(*s)) {
993 do {
994 s++;
995 } while (s < send && isDIGIT(*s));
996 }
3f7602fa
TC
997 else if (flags & PERL_SCAN_TRAILING)
998 return numtype | IS_NUMBER_TRAILING;
60939fb8 999 else
3f7602fa
TC
1000 return 0;
1001
1002 /* The only flag we keep is sign. Blow away any "it's UV" */
1003 numtype &= IS_NUMBER_NEG;
1004 numtype |= IS_NUMBER_NOT_INT;
60939fb8
NC
1005 }
1006 }
1007 while (s < send && isSPACE(*s))
1008 s++;
1009 if (s >= send)
aa8b85de 1010 return numtype;
60939fb8
NC
1011 if (len == 10 && memEQ(pv, "0 but true", 10)) {
1012 if (valuep)
1013 *valuep = 0;
1014 return IS_NUMBER_IN_UV;
1015 }
8c12dc63
JH
1016 /* We could be e.g. at "Inf" or "NaN", or at the "#" of "1.#INF". */
1017 if ((s + 2 < send) && strchr("inqs#", toFOLD(*s))) {
1018 /* Really detect inf/nan. Start at d, not s, since the above
1019 * code might have already consumed the "1." or "1". */
3823048b 1020 int infnan = Perl_grok_infnan(aTHX_ &d, send);
8c12dc63
JH
1021 if ((infnan & IS_NUMBER_INFINITY)) {
1022 return (numtype | infnan); /* Keep sign for infinity. */
1023 }
1024 else if ((infnan & IS_NUMBER_NAN)) {
1025 return (numtype | infnan) & ~IS_NUMBER_NEG; /* Clear sign for nan. */
1026 }
1027 }
3f7602fa
TC
1028 else if (flags & PERL_SCAN_TRAILING) {
1029 return numtype | IS_NUMBER_TRAILING;
1030 }
1031
60939fb8 1032 return 0;
98994639
HS
1033}
1034
6313e544 1035/*
d62b8c6a 1036=for apidoc grok_atou
6313e544 1037
d62b8c6a 1038grok_atou is a safer replacement for atoi and strtol.
6313e544 1039
d62b8c6a
JH
1040grok_atou parses a C-style zero-byte terminated string, looking for
1041a decimal unsigned integer.
338aa8b0 1042
d62b8c6a
JH
1043Returns the unsigned integer, if a valid value can be parsed
1044from the beginning of the string.
f4379102 1045
d62b8c6a 1046Accepts only the decimal digits '0'..'9'.
6313e544 1047
d62b8c6a
JH
1048As opposed to atoi or strtol, grok_atou does NOT allow optional
1049leading whitespace, or negative inputs. If such features are
1050required, the calling code needs to explicitly implement those.
6313e544 1051
d62b8c6a 1052If a valid value cannot be parsed, returns either zero (if non-digits
75feedba 1053are met before any digits) or UV_MAX (if the value overflows).
6313e544 1054
d62b8c6a
JH
1055Note that extraneous leading zeros also count as an overflow
1056(meaning that only "0" is the zero).
338aa8b0 1057
d62b8c6a 1058On failure, the *endptr is also set to NULL, unless endptr is NULL.
338aa8b0
JH
1059
1060Trailing non-digit bytes are allowed if the endptr is non-NULL.
6313e544
JH
1061On return the *endptr will contain the pointer to the first non-digit byte.
1062
6313e544 1063If the endptr is NULL, the first non-digit byte MUST be
f4379102 1064the zero byte terminating the pv, or zero will be returned.
6313e544 1065
d62b8c6a
JH
1066Background: atoi has severe problems with illegal inputs, it cannot be
1067used for incremental parsing, and therefore should be avoided
1068atoi and strtol are also affected by locale settings, which can also be
1069seen as a bug (global state controlled by user environment).
1070
6313e544
JH
1071=cut
1072*/
1073
75feedba 1074UV
6313e544
JH
1075Perl_grok_atou(const char *pv, const char** endptr)
1076{
1077 const char* s = pv;
1078 const char** eptr;
1079 const char* end2; /* Used in case endptr is NULL. */
75feedba 1080 UV val = 0; /* The return value. */
6313e544
JH
1081
1082 PERL_ARGS_ASSERT_GROK_ATOU;
1083
1084 eptr = endptr ? endptr : &end2;
75feedba
JH
1085 if (isDIGIT(*s)) {
1086 /* Single-digit inputs are quite common. */
6313e544 1087 val = *s++ - '0';
75feedba
JH
1088 if (isDIGIT(*s)) {
1089 /* Extra leading zeros cause overflow. */
1090 if (val == 0) {
1091 *eptr = NULL;
1092 return UV_MAX;
1093 }
1094 while (isDIGIT(*s)) {
1095 /* This could be unrolled like in grok_number(), but
1096 * the expected uses of this are not speed-needy, and
1097 * unlikely to need full 64-bitness. */
1098 U8 digit = *s++ - '0';
945b524a
JH
1099 if (val < uv_max_div_10 ||
1100 (val == uv_max_div_10 && digit <= uv_max_mod_10)) {
75feedba
JH
1101 val = val * 10 + digit;
1102 } else {
6313e544 1103 *eptr = NULL;
75feedba 1104 return UV_MAX;
6313e544 1105 }
6313e544
JH
1106 }
1107 }
75feedba
JH
1108 }
1109 if (s == pv) {
1110 *eptr = NULL; /* If no progress, failed to parse anything. */
1111 return 0;
6313e544
JH
1112 }
1113 if (endptr == NULL && *s) {
1114 return 0; /* If endptr is NULL, no trailing non-digits allowed. */
1115 }
1116 *eptr = s;
1117 return val;
1118}
1119
a4eca1d4 1120#ifndef USE_QUADMATH
4801ca72 1121STATIC NV
98994639
HS
1122S_mulexp10(NV value, I32 exponent)
1123{
1124 NV result = 1.0;
1125 NV power = 10.0;
1126 bool negative = 0;
1127 I32 bit;
1128
1129 if (exponent == 0)
1130 return value;
659c4b96
DM
1131 if (value == 0)
1132 return (NV)0;
87032ba1 1133
24866caa 1134 /* On OpenVMS VAX we by default use the D_FLOAT double format,
67597c89 1135 * and that format does not have *easy* capabilities [1] for
24866caa
CB
1136 * overflowing doubles 'silently' as IEEE fp does. We also need
1137 * to support G_FLOAT on both VAX and Alpha, and though the exponent
1138 * range is much larger than D_FLOAT it still doesn't do silent
1139 * overflow. Therefore we need to detect early whether we would
1140 * overflow (this is the behaviour of the native string-to-float
1141 * conversion routines, and therefore of native applications, too).
67597c89 1142 *
24866caa
CB
1143 * [1] Trying to establish a condition handler to trap floating point
1144 * exceptions is not a good idea. */
87032ba1
JH
1145
1146 /* In UNICOS and in certain Cray models (such as T90) there is no
1147 * IEEE fp, and no way at all from C to catch fp overflows gracefully.
1148 * There is something you can do if you are willing to use some
1149 * inline assembler: the instruction is called DFI-- but that will
1150 * disable *all* floating point interrupts, a little bit too large
1151 * a hammer. Therefore we need to catch potential overflows before
1152 * it's too late. */
353813d9 1153
85bba25f 1154#if ((defined(VMS) && !defined(_IEEE_FP)) || defined(_UNICOS)) && defined(NV_MAX_10_EXP)
353813d9 1155 STMT_START {
c4420975 1156 const NV exp_v = log10(value);
353813d9
HS
1157 if (exponent >= NV_MAX_10_EXP || exponent + exp_v >= NV_MAX_10_EXP)
1158 return NV_MAX;
1159 if (exponent < 0) {
1160 if (-(exponent + exp_v) >= NV_MAX_10_EXP)
1161 return 0.0;
1162 while (-exponent >= NV_MAX_10_EXP) {
1163 /* combination does not overflow, but 10^(-exponent) does */
1164 value /= 10;
1165 ++exponent;
1166 }
1167 }
1168 } STMT_END;
87032ba1
JH
1169#endif
1170
353813d9
HS
1171 if (exponent < 0) {
1172 negative = 1;
1173 exponent = -exponent;
b27804d8
DM
1174#ifdef NV_MAX_10_EXP
1175 /* for something like 1234 x 10^-309, the action of calculating
1176 * the intermediate value 10^309 then returning 1234 / (10^309)
1177 * will fail, since 10^309 becomes infinity. In this case try to
1178 * refactor it as 123 / (10^308) etc.
1179 */
1180 while (value && exponent > NV_MAX_10_EXP) {
1181 exponent--;
1182 value /= 10;
1183 }
48853916
JH
1184 if (value == 0.0)
1185 return value;
b27804d8 1186#endif
353813d9 1187 }
c62e754c
JH
1188#if defined(__osf__)
1189 /* Even with cc -ieee + ieee_set_fp_control(IEEE_TRAP_ENABLE_INV)
1190 * Tru64 fp behavior on inf/nan is somewhat broken. Another way
1191 * to do this would be ieee_set_fp_control(IEEE_TRAP_ENABLE_OVF)
1192 * but that breaks another set of infnan.t tests. */
1193# define FP_OVERFLOWS_TO_ZERO
1194#endif
98994639
HS
1195 for (bit = 1; exponent; bit <<= 1) {
1196 if (exponent & bit) {
1197 exponent ^= bit;
1198 result *= power;
c62e754c
JH
1199#ifdef FP_OVERFLOWS_TO_ZERO
1200 if (result == 0)
1201 return value < 0 ? -NV_INF : NV_INF;
1202#endif
236f0012
CB
1203 /* Floating point exceptions are supposed to be turned off,
1204 * but if we're obviously done, don't risk another iteration.
1205 */
1206 if (exponent == 0) break;
98994639
HS
1207 }
1208 power *= power;
1209 }
1210 return negative ? value / result : value * result;
1211}
a4eca1d4 1212#endif /* #ifndef USE_QUADMATH */
98994639
HS
1213
1214NV
1215Perl_my_atof(pTHX_ const char* s)
1216{
1217 NV x = 0.0;
a4eca1d4
JH
1218#ifdef USE_QUADMATH
1219 Perl_my_atof2(aTHX_ s, &x);
1220 return x;
1221#else
1222# ifdef USE_LOCALE_NUMERIC
7918f24d
NC
1223 PERL_ARGS_ASSERT_MY_ATOF;
1224
a2287a13
KW
1225 {
1226 DECLARE_STORE_LC_NUMERIC_SET_TO_NEEDED();
d6ded950 1227 if (PL_numeric_radix_sv && IN_LC(LC_NUMERIC)) {
e4850248
KW
1228 const char *standard = NULL, *local = NULL;
1229 bool use_standard_radix;
98994639 1230
e4850248
KW
1231 /* Look through the string for the first thing that looks like a
1232 * decimal point: either the value in the current locale or the
1233 * standard fallback of '.'. The one which appears earliest in the
1234 * input string is the one that we should have atof look for. Note
1235 * that we have to determine this beforehand because on some
1236 * systems, Perl_atof2 is just a wrapper around the system's atof.
1237 * */
1238 standard = strchr(s, '.');
1239 local = strstr(s, SvPV_nolen(PL_numeric_radix_sv));
78787052 1240
e4850248 1241 use_standard_radix = standard && (!local || standard < local);
78787052 1242
e4850248
KW
1243 if (use_standard_radix)
1244 SET_NUMERIC_STANDARD();
78787052 1245
e4850248 1246 Perl_atof2(s, x);
78787052 1247
e4850248
KW
1248 if (use_standard_radix)
1249 SET_NUMERIC_LOCAL();
1250 }
1251 else
1252 Perl_atof2(s, x);
a2287a13
KW
1253 RESTORE_LC_NUMERIC();
1254 }
a4eca1d4 1255# else
a36244b7 1256 Perl_atof2(s, x);
a4eca1d4 1257# endif
98994639
HS
1258#endif
1259 return x;
1260}
1261
3c81f0b3
DD
1262
1263#ifdef USING_MSVC6
1264# pragma warning(push)
1265# pragma warning(disable:4756;disable:4056)
1266#endif
829757a4 1267static char*
5563f457 1268S_my_atof_infnan(pTHX_ const char* s, bool negative, const char* send, NV* value)
829757a4
JH
1269{
1270 const char *p0 = negative ? s - 1 : s;
1271 const char *p = p0;
3823048b 1272 int infnan = grok_infnan(&p, send);
829757a4
JH
1273 if (infnan && p != p0) {
1274 /* If we can generate inf/nan directly, let's do so. */
1275#ifdef NV_INF
1276 if ((infnan & IS_NUMBER_INFINITY)) {
3823048b 1277 *value = (infnan & IS_NUMBER_NEG) ? -NV_INF: NV_INF;
829757a4
JH
1278 return (char*)p;
1279 }
1280#endif
1281#ifdef NV_NAN
1282 if ((infnan & IS_NUMBER_NAN)) {
3823048b 1283 *value = NV_NAN;
829757a4
JH
1284 return (char*)p;
1285 }
1286#endif
1287#ifdef Perl_strtod
68611e6f 1288 /* If still here, we didn't have either NV_INF or NV_NAN,
829757a4
JH
1289 * and can try falling back to native strtod/strtold.
1290 *
68611e6f
JH
1291 * (Though, are our NV_INF or NV_NAN ever not defined?)
1292 *
829757a4
JH
1293 * The native interface might not recognize all the possible
1294 * inf/nan strings Perl recognizes. What we can try
1295 * is to try faking the input. We will try inf/-inf/nan
1296 * as the most promising/portable input. */
1297 {
1298 const char* fake = NULL;
1299 char* endp;
1300 NV nv;
1301 if ((infnan & IS_NUMBER_INFINITY)) {
1302 fake = ((infnan & IS_NUMBER_NEG)) ? "-inf" : "inf";
1303 }
1304 else if ((infnan & IS_NUMBER_NAN)) {
1305 fake = "nan";
1306 }
1307 assert(fake);
1308 nv = Perl_strtod(fake, &endp);
1309 if (fake != endp) {
1310 if ((infnan & IS_NUMBER_INFINITY)) {
1311#ifdef Perl_isinf
1312 if (Perl_isinf(nv))
1313 *value = nv;
1314#else
1315 /* last resort, may generate SIGFPE */
1316 *value = Perl_exp((NV)1e9);
1317 if ((infnan & IS_NUMBER_NEG))
1318 *value = -*value;
1319#endif
1320 return (char*)p; /* p, not endp */
1321 }
1322 else if ((infnan & IS_NUMBER_NAN)) {
1323#ifdef Perl_isnan
1324 if (Perl_isnan(nv))
1325 *value = nv;
1326#else
1327 /* last resort, may generate SIGFPE */
1328 *value = Perl_log((NV)-1.0);
1329#endif
1330 return (char*)p; /* p, not endp */
1331 }
1332 }
1333 }
1334#endif /* #ifdef Perl_strtod */
1335 }
1336 return NULL;
1337}
3c81f0b3
DD
1338#ifdef USING_MSVC6
1339# pragma warning(pop)
1340#endif
829757a4 1341
98994639
HS
1342char*
1343Perl_my_atof2(pTHX_ const char* orig, NV* value)
1344{
e1ec3a88 1345 const char* s = orig;
a4eca1d4
JH
1346 NV result[3] = {0.0, 0.0, 0.0};
1347#if defined(USE_PERL_ATOF) || defined(USE_QUADMATH)
ae776a2c 1348 const char* send = s + strlen(orig); /* one past the last */
a4eca1d4
JH
1349 bool negative = 0;
1350#endif
1351#if defined(USE_PERL_ATOF) && !defined(USE_QUADMATH)
1352 UV accumulator[2] = {0,0}; /* before/after dp */
8194bf88 1353 bool seen_digit = 0;
20f6aaab
AS
1354 I32 exp_adjust[2] = {0,0};
1355 I32 exp_acc[2] = {-1, -1};
1356 /* the current exponent adjust for the accumulators */
98994639 1357 I32 exponent = 0;
8194bf88 1358 I32 seen_dp = 0;
20f6aaab
AS
1359 I32 digit = 0;
1360 I32 old_digit = 0;
8194bf88 1361 I32 sig_digits = 0; /* noof significant digits seen so far */
a4eca1d4 1362#endif
8194bf88 1363
a4eca1d4 1364#if defined(USE_PERL_ATOF) || defined(USE_QUADMATH)
7918f24d
NC
1365 PERL_ARGS_ASSERT_MY_ATOF2;
1366
a4eca1d4
JH
1367 /* leading whitespace */
1368 while (isSPACE(*s))
1369 ++s;
1370
1371 /* sign */
1372 switch (*s) {
1373 case '-':
1374 negative = 1;
1375 /* FALLTHROUGH */
1376 case '+':
1377 ++s;
1378 }
1379#endif
1380
1381#ifdef USE_QUADMATH
1382 {
1383 char* endp;
adc55e02 1384 if ((endp = S_my_atof_infnan(aTHX_ s, negative, send, value)))
a4eca1d4
JH
1385 return endp;
1386 result[2] = strtoflt128(s, &endp);
1387 if (s != endp) {
1388 *value = negative ? -result[2] : result[2];
1389 return endp;
1390 }
1391 return NULL;
1392 }
1393#elif defined(USE_PERL_ATOF)
1394
8194bf88
DM
1395/* There is no point in processing more significant digits
1396 * than the NV can hold. Note that NV_DIG is a lower-bound value,
1397 * while we need an upper-bound value. We add 2 to account for this;
1398 * since it will have been conservative on both the first and last digit.
1399 * For example a 32-bit mantissa with an exponent of 4 would have
1400 * exact values in the set
1401 * 4
1402 * 8
1403 * ..
1404 * 17179869172
1405 * 17179869176
1406 * 17179869180
1407 *
1408 * where for the purposes of calculating NV_DIG we would have to discount
1409 * both the first and last digit, since neither can hold all values from
1410 * 0..9; but for calculating the value we must examine those two digits.
1411 */
ffa277e5
EAS
1412#ifdef MAX_SIG_DIG_PLUS
1413 /* It is not necessarily the case that adding 2 to NV_DIG gets all the
1414 possible digits in a NV, especially if NVs are not IEEE compliant
1415 (e.g., long doubles on IRIX) - Allen <allens@cpan.org> */
1416# define MAX_SIG_DIGITS (NV_DIG+MAX_SIG_DIG_PLUS)
1417#else
1418# define MAX_SIG_DIGITS (NV_DIG+2)
1419#endif
8194bf88
DM
1420
1421/* the max number we can accumulate in a UV, and still safely do 10*N+9 */
1422#define MAX_ACCUMULATE ( (UV) ((UV_MAX - 9)/10))
98994639 1423
ae776a2c 1424 {
829757a4 1425 const char* endp;
5563f457 1426 if ((endp = S_my_atof_infnan(aTHX_ s, negative, send, value)))
829757a4 1427 return (char*)endp;
ae776a2c 1428 }
2b54f59f 1429
8194bf88
DM
1430 /* we accumulate digits into an integer; when this becomes too
1431 * large, we add the total to NV and start again */
98994639 1432
8194bf88
DM
1433 while (1) {
1434 if (isDIGIT(*s)) {
1435 seen_digit = 1;
20f6aaab 1436 old_digit = digit;
8194bf88 1437 digit = *s++ - '0';
20f6aaab
AS
1438 if (seen_dp)
1439 exp_adjust[1]++;
98994639 1440
8194bf88
DM
1441 /* don't start counting until we see the first significant
1442 * digit, eg the 5 in 0.00005... */
1443 if (!sig_digits && digit == 0)
1444 continue;
1445
1446 if (++sig_digits > MAX_SIG_DIGITS) {
98994639 1447 /* limits of precision reached */
20f6aaab
AS
1448 if (digit > 5) {
1449 ++accumulator[seen_dp];
1450 } else if (digit == 5) {
1451 if (old_digit % 2) { /* round to even - Allen */
1452 ++accumulator[seen_dp];
1453 }
1454 }
1455 if (seen_dp) {
1456 exp_adjust[1]--;
1457 } else {
1458 exp_adjust[0]++;
1459 }
8194bf88 1460 /* skip remaining digits */
98994639 1461 while (isDIGIT(*s)) {
98994639 1462 ++s;
20f6aaab
AS
1463 if (! seen_dp) {
1464 exp_adjust[0]++;
1465 }
98994639
HS
1466 }
1467 /* warn of loss of precision? */
98994639 1468 }
8194bf88 1469 else {
20f6aaab 1470 if (accumulator[seen_dp] > MAX_ACCUMULATE) {
8194bf88 1471 /* add accumulator to result and start again */
20f6aaab
AS
1472 result[seen_dp] = S_mulexp10(result[seen_dp],
1473 exp_acc[seen_dp])
1474 + (NV)accumulator[seen_dp];
1475 accumulator[seen_dp] = 0;
1476 exp_acc[seen_dp] = 0;
98994639 1477 }
20f6aaab
AS
1478 accumulator[seen_dp] = accumulator[seen_dp] * 10 + digit;
1479 ++exp_acc[seen_dp];
98994639 1480 }
8194bf88 1481 }
e1ec3a88 1482 else if (!seen_dp && GROK_NUMERIC_RADIX(&s, send)) {
8194bf88 1483 seen_dp = 1;
20f6aaab 1484 if (sig_digits > MAX_SIG_DIGITS) {
c86f7df5 1485 do {
20f6aaab 1486 ++s;
c86f7df5 1487 } while (isDIGIT(*s));
20f6aaab
AS
1488 break;
1489 }
8194bf88
DM
1490 }
1491 else {
1492 break;
98994639
HS
1493 }
1494 }
1495
20f6aaab
AS
1496 result[0] = S_mulexp10(result[0], exp_acc[0]) + (NV)accumulator[0];
1497 if (seen_dp) {
1498 result[1] = S_mulexp10(result[1], exp_acc[1]) + (NV)accumulator[1];
1499 }
98994639 1500
305b8651 1501 if (seen_digit && (isALPHA_FOLD_EQ(*s, 'e'))) {
98994639
HS
1502 bool expnegative = 0;
1503
1504 ++s;
1505 switch (*s) {
1506 case '-':
1507 expnegative = 1;
924ba076 1508 /* FALLTHROUGH */
98994639
HS
1509 case '+':
1510 ++s;
1511 }
1512 while (isDIGIT(*s))
1513 exponent = exponent * 10 + (*s++ - '0');
1514 if (expnegative)
1515 exponent = -exponent;
1516 }
1517
20f6aaab
AS
1518
1519
98994639 1520 /* now apply the exponent */
20f6aaab
AS
1521
1522 if (seen_dp) {
1523 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0])
1524 + S_mulexp10(result[1],exponent-exp_adjust[1]);
1525 } else {
1526 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0]);
1527 }
98994639
HS
1528
1529 /* now apply the sign */
1530 if (negative)
20f6aaab 1531 result[2] = -result[2];
a36244b7 1532#endif /* USE_PERL_ATOF */
20f6aaab 1533 *value = result[2];
73d840c0 1534 return (char *)s;
98994639
HS
1535}
1536
5d34af89 1537/*
3d9d9213 1538=for apidoc isinfnan
5d34af89
JH
1539
1540Perl_isinfnan() is utility function that returns true if the NV
1541argument is either an infinity or a NaN, false otherwise. To test
1542in more detail, use Perl_isinf() and Perl_isnan().
1543
68611e6f
JH
1544This is also the logical inverse of Perl_isfinite().
1545
5d34af89
JH
1546=cut
1547*/
1cd88304
JH
1548bool
1549Perl_isinfnan(NV nv)
1550{
1551#ifdef Perl_isinf
1552 if (Perl_isinf(nv))
1553 return TRUE;
1554#endif
1555#ifdef Perl_isnan
1556 if (Perl_isnan(nv))
1557 return TRUE;
1558#endif
1559 return FALSE;
1560}
1561
354b74ae
FC
1562/*
1563=for apidoc
1564
1565Checks whether the argument would be either an infinity or NaN when used
1566as a number, but is careful not to trigger non-numeric or uninitialized
1567warnings. it assumes the caller has done SvGETMAGIC(sv) already.
1568
1569=cut
1570*/
1571
1572bool
1573Perl_isinfnansv(pTHX_ SV *sv)
1574{
1575 PERL_ARGS_ASSERT_ISINFNANSV;
1576 if (!SvOK(sv))
1577 return FALSE;
1578 if (SvNOKp(sv))
1579 return Perl_isinfnan(SvNVX(sv));
1580 if (SvIOKp(sv))
1581 return FALSE;
1582 {
1583 STRLEN len;
1584 const char *s = SvPV_nomg_const(sv, len);
3823048b 1585 return cBOOL(grok_infnan(&s, s+len));
354b74ae
FC
1586 }
1587}
1588
d67dac15 1589#ifndef HAS_MODFL
68611e6f
JH
1590/* C99 has truncl, pre-C99 Solaris had aintl. We can use either with
1591 * copysignl to emulate modfl, which is in some platforms missing or
1592 * broken. */
d67dac15
JH
1593# if defined(HAS_TRUNCL) && defined(HAS_COPYSIGNL)
1594long double
1595Perl_my_modfl(long double x, long double *ip)
1596{
68611e6f
JH
1597 *ip = truncl(x);
1598 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
d67dac15
JH
1599}
1600# elif defined(HAS_AINTL) && defined(HAS_COPYSIGNL)
55954f19
JH
1601long double
1602Perl_my_modfl(long double x, long double *ip)
1603{
68611e6f
JH
1604 *ip = aintl(x);
1605 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
55954f19 1606}
d67dac15 1607# endif
55954f19
JH
1608#endif
1609
7b9b7dff 1610/* Similarly, with ilogbl and scalbnl we can emulate frexpl. */
55954f19
JH
1611#if ! defined(HAS_FREXPL) && defined(HAS_ILOGBL) && defined(HAS_SCALBNL)
1612long double
1613Perl_my_frexpl(long double x, int *e) {
68611e6f
JH
1614 *e = x == 0.0L ? 0 : ilogbl(x) + 1;
1615 return (scalbnl(x, -*e));
55954f19
JH
1616}
1617#endif
66610fdd
RGS
1618
1619/*
ed140128
AD
1620=for apidoc Perl_signbit
1621
1622Return a non-zero integer if the sign bit on an NV is set, and 0 if
1623it is not.
1624
1625If Configure detects this system has a signbit() that will work with
1626our NVs, then we just use it via the #define in perl.h. Otherwise,
8b7fad81
JH
1627fall back on this implementation. The main use of this function
1628is catching -0.0.
ed140128
AD
1629
1630Configure notes: This function is called 'Perl_signbit' instead of a
1631plain 'signbit' because it is easy to imagine a system having a signbit()
1632function or macro that doesn't happen to work with our particular choice
1633of NVs. We shouldn't just re-#define signbit as Perl_signbit and expect
1634the standard system headers to be happy. Also, this is a no-context
1635function (no pTHX_) because Perl_signbit() is usually re-#defined in
1636perl.h as a simple macro call to the system's signbit().
1637Users should just always call Perl_signbit().
1638
1639=cut
1640*/
1641#if !defined(HAS_SIGNBIT)
1642int
1643Perl_signbit(NV x) {
8b7fad81
JH
1644# ifdef Perl_fp_class_nzero
1645 if (x == 0)
1646 return Perl_fp_class_nzero(x);
8b7fad81 1647# endif
3585840c 1648 return (x < 0.0) ? 1 : 0;
ed140128
AD
1649}
1650#endif
1651
1652/*
66610fdd
RGS
1653 * Local variables:
1654 * c-indentation-style: bsd
1655 * c-basic-offset: 4
14d04a33 1656 * indent-tabs-mode: nil
66610fdd
RGS
1657 * End:
1658 *
14d04a33 1659 * ex: set ts=8 sts=4 sw=4 et:
37442d52 1660 */