This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
infnan: comment tweaks
[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/*
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
5563f457 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++;
3396ed30
JH
609 while (*s == '0') { /* 1.#INF00 */
610 s++;
611 }
b489e20f
JH
612 while (s < send && isSPACE(*s))
613 s++;
614 if (s < send && *s) {
3396ed30 615 flags |= IS_NUMBER_TRAILING;
fae4db12 616 }
ff4eb398
JH
617 flags |= IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT;
618 }
e855f543 619 else if (isALPHA_FOLD_EQ(*s, 'D') && odh) { /* 1.#IND */
ff4eb398
JH
620 s++;
621 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
fae4db12
JH
622 while (*s == '0') { /* 1.#IND00 */
623 s++;
624 }
1e9aa12f
JH
625 if (*s) {
626 flags |= IS_NUMBER_TRAILING;
627 }
ff4eb398
JH
628 } else
629 return 0;
ff4eb398
JH
630 }
631 else {
62bdc035
JH
632 /* Maybe NAN of some sort */
633
305b8651 634 if (isALPHA_FOLD_EQ(*s, 'S') || isALPHA_FOLD_EQ(*s, 'Q')) {
ff4eb398
JH
635 /* snan, qNaN */
636 /* XXX do something with the snan/qnan difference */
ae776a2c 637 s++; if (s == send) return 0;
ff4eb398
JH
638 }
639
305b8651
KW
640 if (isALPHA_FOLD_EQ(*s, 'N')) {
641 s++; if (s == send || isALPHA_FOLD_NE(*s, 'A')) return 0;
642 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
ff4eb398
JH
643 s++;
644
ae776a2c
JH
645 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
646
8c12dc63
JH
647 /* NaN can be followed by various stuff (NaNQ, NaNS), but
648 * there are also multiple different NaN values, and some
649 * implementations output the "payload" values,
1e9aa12f 650 * e.g. NaN123, NAN(abc), while some legacy implementations
8c12dc63 651 * have weird stuff like NaN%. */
1e9aa12f
JH
652 if (isALPHA_FOLD_EQ(*s, 'q') ||
653 isALPHA_FOLD_EQ(*s, 's')) {
654 /* "nanq" or "nans" are ok, though generating
655 * these portably is tricky. */
656 s++;
657 }
658 if (*s == '(') {
659 /* C99 style "nan(123)" or Perlish equivalent "nan($uv)". */
660 const char *t;
1e9aa12f
JH
661 s++;
662 if (s == send) {
663 return flags | IS_NUMBER_TRAILING;
664 }
665 t = s + 1;
666 while (t < send && *t && *t != ')') {
667 t++;
668 }
669 if (t == send) {
670 return flags | IS_NUMBER_TRAILING;
671 }
672 if (*t == ')') {
bf8c8f7f
JH
673 int nantype;
674 UV nanval;
675 if (s[0] == '0' && s + 2 < t &&
676 isALPHA_FOLD_EQ(s[1], 'x') &&
677 isXDIGIT(s[2])) {
678 STRLEN len = t - s;
679 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES;
680 nanval = grok_hex(s, &len, &flags, NULL);
681 if ((flags & PERL_SCAN_GREATER_THAN_UV_MAX)) {
682 nantype = 0;
683 } else {
684 nantype = IS_NUMBER_IN_UV;
685 }
686 s += len;
687 } else if (s[0] == '0' && s + 2 < t &&
688 isALPHA_FOLD_EQ(s[1], 'b') &&
689 (s[2] == '0' || s[2] == '1')) {
690 STRLEN len = t - s;
691 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES;
692 nanval = grok_bin(s, &len, &flags, NULL);
693 if ((flags & PERL_SCAN_GREATER_THAN_UV_MAX)) {
694 nantype = 0;
695 } else {
696 nantype = IS_NUMBER_IN_UV;
697 }
698 s += len;
699 } else {
700 const char *u;
701 nantype =
702 grok_number_flags(s, t - s, &nanval,
703 PERL_SCAN_TRAILING |
704 PERL_SCAN_ALLOW_UNDERSCORES);
705 /* Unfortunately grok_number_flags() doesn't
706 * tell how far we got and the ')' will always
707 * be "trailing", so we need to double-check
708 * whether we had something dubious. */
709 for (u = s; u < t; u++) {
710 if (!isDIGIT(*u)) {
711 flags |= IS_NUMBER_TRAILING;
712 break;
713 }
714 }
715 s = u;
716 }
717
718 /* "What about octal?" Really? */
719
1e9aa12f
JH
720 if ((nantype & IS_NUMBER_NOT_INT) ||
721 !(nantype && IS_NUMBER_IN_UV)) {
722 /* Certain configuration combinations where
723 * NVSIZE is greater than UVSIZE mean that a
724 * single UV cannot contain all the possible
725 * NaN payload bits. There would need to be
726 * some more generic syntax than "nan($uv)".
727 * Issues to keep in mind:
728 * (1) In most common cases there would
729 * not be an integral number of bytes that
730 * could be set, only a certain number of bits.
731 * For example for NVSIZE==UVSIZE there can be
732 * up to 52 bits in the payload, but one bit is
733 * commonly reserved for the signal/quiet bit,
734 * so 51 bits.
735 * (2) Endianness of the payload bits. If the
736 * payload is specified as an UV, the low-order
737 * bits of the UV are naturally little-endianed
738 * (rightmost) bits of the payload. */
739 return 0;
740 }
1e9aa12f
JH
741 if (s < t) {
742 flags |= IS_NUMBER_TRAILING;
743 }
744 } else {
745 /* Looked like nan(...), but no close paren. */
746 flags |= IS_NUMBER_TRAILING;
747 }
b489e20f
JH
748 } else {
749 while (s < send && isSPACE(*s))
750 s++;
751 if (s < send && *s) {
752 /* Note that we here implicitly accept (parse as
753 * "nan", but with warnings) also any other weird
754 * trailing stuff for "nan". In the above we just
755 * check that if we got the C99-style "nan(...)",
756 * the "..." looks sane.
757 * If in future we accept more ways of specifying
758 * the nan payload, the accepting would happen around
759 * here. */
760 flags |= IS_NUMBER_TRAILING;
761 }
1e9aa12f 762 }
ae776a2c 763 s = send;
ff4eb398 764 }
ae776a2c
JH
765 else
766 return 0;
ff4eb398
JH
767 }
768
b489e20f
JH
769 while (s < send && isSPACE(*s))
770 s++;
771
a1fe7cea
JH
772 *sp = s;
773 return flags;
ff4eb398
JH
774}
775
13393a5e
JH
776/*
777=for apidoc grok_number_flags
778
779Recognise (or not) a number. The type of the number is returned
780(0 if unrecognised), otherwise it is a bit-ORed combination of
781IS_NUMBER_IN_UV, IS_NUMBER_GREATER_THAN_UV_MAX, IS_NUMBER_NOT_INT,
782IS_NUMBER_NEG, IS_NUMBER_INFINITY, IS_NUMBER_NAN (defined in perl.h).
783
784If the value of the number can fit in a UV, it is returned in the *valuep
785IS_NUMBER_IN_UV will be set to indicate that *valuep is valid, IS_NUMBER_IN_UV
786will never be set unless *valuep is valid, but *valuep may have been assigned
787to during processing even though IS_NUMBER_IN_UV is not set on return.
788If valuep is NULL, IS_NUMBER_IN_UV will be set for the same cases as when
789valuep is non-NULL, but no actual assignment (or SEGV) will occur.
790
791IS_NUMBER_NOT_INT will be set with IS_NUMBER_IN_UV if trailing decimals were
792seen (in which case *valuep gives the true value truncated to an integer), and
793IS_NUMBER_NEG if the number is negative (in which case *valuep holds the
794absolute value). IS_NUMBER_IN_UV is not set if e notation was used or the
795number is larger than a UV.
796
797C<flags> allows only C<PERL_SCAN_TRAILING>, which allows for trailing
798non-numeric text on an otherwise successful I<grok>, setting
799C<IS_NUMBER_TRAILING> on the result.
800
801=for apidoc grok_number
802
803Identical to grok_number_flags() with flags set to zero.
804
805=cut
806 */
807int
808Perl_grok_number(pTHX_ const char *pv, STRLEN len, UV *valuep)
809{
810 PERL_ARGS_ASSERT_GROK_NUMBER;
811
812 return grok_number_flags(pv, len, valuep, 0);
813}
814
945b524a
JH
815static const UV uv_max_div_10 = UV_MAX / 10;
816static const U8 uv_max_mod_10 = UV_MAX % 10;
817
3f7602fa
TC
818int
819Perl_grok_number_flags(pTHX_ const char *pv, STRLEN len, UV *valuep, U32 flags)
820{
60939fb8 821 const char *s = pv;
c4420975 822 const char * const send = pv + len;
ae776a2c 823 const char *d;
60939fb8 824 int numtype = 0;
60939fb8 825
3f7602fa 826 PERL_ARGS_ASSERT_GROK_NUMBER_FLAGS;
7918f24d 827
60939fb8
NC
828 while (s < send && isSPACE(*s))
829 s++;
830 if (s == send) {
831 return 0;
832 } else if (*s == '-') {
833 s++;
834 numtype = IS_NUMBER_NEG;
835 }
836 else if (*s == '+')
aa42a541 837 s++;
60939fb8
NC
838
839 if (s == send)
840 return 0;
841
ae776a2c 842 /* The first digit (after optional sign): note that might
8c12dc63 843 * also point to "infinity" or "nan", or "1.#INF". */
ae776a2c
JH
844 d = s;
845
8c12dc63 846 /* next must be digit or the radix separator or beginning of infinity/nan */
60939fb8
NC
847 if (isDIGIT(*s)) {
848 /* UVs are at least 32 bits, so the first 9 decimal digits cannot
849 overflow. */
850 UV value = *s - '0';
851 /* This construction seems to be more optimiser friendly.
852 (without it gcc does the isDIGIT test and the *s - '0' separately)
853 With it gcc on arm is managing 6 instructions (6 cycles) per digit.
854 In theory the optimiser could deduce how far to unroll the loop
855 before checking for overflow. */
58bb9ec3
NC
856 if (++s < send) {
857 int 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
NC
868 if (++s < send) {
869 digit = *s - '0';
60939fb8
NC
870 if (digit >= 0 && digit <= 9) {
871 value = value * 10 + digit;
58bb9ec3
NC
872 if (++s < send) {
873 digit = *s - '0';
60939fb8
NC
874 if (digit >= 0 && digit <= 9) {
875 value = value * 10 + digit;
58bb9ec3
NC
876 if (++s < send) {
877 digit = *s - '0';
60939fb8
NC
878 if (digit >= 0 && digit <= 9) {
879 value = value * 10 + digit;
58bb9ec3
NC
880 if (++s < send) {
881 digit = *s - '0';
60939fb8
NC
882 if (digit >= 0 && digit <= 9) {
883 value = value * 10 + digit;
58bb9ec3
NC
884 if (++s < send) {
885 digit = *s - '0';
60939fb8
NC
886 if (digit >= 0 && digit <= 9) {
887 value = value * 10 + digit;
58bb9ec3 888 if (++s < send) {
60939fb8
NC
889 /* Now got 9 digits, so need to check
890 each time for overflow. */
58bb9ec3 891 digit = *s - '0';
60939fb8 892 while (digit >= 0 && digit <= 9
945b524a
JH
893 && (value < uv_max_div_10
894 || (value == uv_max_div_10
895 && digit <= uv_max_mod_10))) {
60939fb8 896 value = value * 10 + digit;
58bb9ec3
NC
897 if (++s < send)
898 digit = *s - '0';
60939fb8
NC
899 else
900 break;
901 }
902 if (digit >= 0 && digit <= 9
51bd16da 903 && (s < send)) {
60939fb8
NC
904 /* value overflowed.
905 skip the remaining digits, don't
906 worry about setting *valuep. */
907 do {
908 s++;
909 } while (s < send && isDIGIT(*s));
910 numtype |=
911 IS_NUMBER_GREATER_THAN_UV_MAX;
912 goto skip_value;
913 }
914 }
915 }
98994639 916 }
60939fb8
NC
917 }
918 }
919 }
920 }
921 }
922 }
923 }
924 }
925 }
926 }
927 }
98994639 928 }
60939fb8 929 }
98994639 930 }
60939fb8
NC
931 numtype |= IS_NUMBER_IN_UV;
932 if (valuep)
933 *valuep = value;
934
935 skip_value:
936 if (GROK_NUMERIC_RADIX(&s, send)) {
937 numtype |= IS_NUMBER_NOT_INT;
938 while (s < send && isDIGIT(*s)) /* optional digits after the radix */
939 s++;
98994639 940 }
60939fb8
NC
941 }
942 else if (GROK_NUMERIC_RADIX(&s, send)) {
943 numtype |= IS_NUMBER_NOT_INT | IS_NUMBER_IN_UV; /* valuep assigned below */
944 /* no digits before the radix means we need digits after it */
945 if (s < send && isDIGIT(*s)) {
946 do {
947 s++;
948 } while (s < send && isDIGIT(*s));
949 if (valuep) {
950 /* integer approximation is valid - it's 0. */
951 *valuep = 0;
952 }
98994639 953 }
60939fb8 954 else
ae776a2c 955 return 0;
ff4eb398 956 }
60939fb8 957
926f5fc6 958 if (s > d && s < send) {
60939fb8 959 /* we can have an optional exponent part */
305b8651 960 if (isALPHA_FOLD_EQ(*s, 'e')) {
60939fb8
NC
961 s++;
962 if (s < send && (*s == '-' || *s == '+'))
963 s++;
964 if (s < send && isDIGIT(*s)) {
965 do {
966 s++;
967 } while (s < send && isDIGIT(*s));
968 }
3f7602fa
TC
969 else if (flags & PERL_SCAN_TRAILING)
970 return numtype | IS_NUMBER_TRAILING;
60939fb8 971 else
3f7602fa
TC
972 return 0;
973
974 /* The only flag we keep is sign. Blow away any "it's UV" */
975 numtype &= IS_NUMBER_NEG;
976 numtype |= IS_NUMBER_NOT_INT;
60939fb8
NC
977 }
978 }
979 while (s < send && isSPACE(*s))
980 s++;
981 if (s >= send)
aa8b85de 982 return numtype;
60939fb8
NC
983 if (len == 10 && memEQ(pv, "0 but true", 10)) {
984 if (valuep)
985 *valuep = 0;
986 return IS_NUMBER_IN_UV;
987 }
8c12dc63
JH
988 /* We could be e.g. at "Inf" or "NaN", or at the "#" of "1.#INF". */
989 if ((s + 2 < send) && strchr("inqs#", toFOLD(*s))) {
990 /* Really detect inf/nan. Start at d, not s, since the above
991 * code might have already consumed the "1." or "1". */
5563f457 992 int infnan = Perl_grok_infnan(aTHX_ &d, send);
8c12dc63
JH
993 if ((infnan & IS_NUMBER_INFINITY)) {
994 return (numtype | infnan); /* Keep sign for infinity. */
995 }
996 else if ((infnan & IS_NUMBER_NAN)) {
997 return (numtype | infnan) & ~IS_NUMBER_NEG; /* Clear sign for nan. */
998 }
999 }
3f7602fa
TC
1000 else if (flags & PERL_SCAN_TRAILING) {
1001 return numtype | IS_NUMBER_TRAILING;
1002 }
1003
60939fb8 1004 return 0;
98994639
HS
1005}
1006
6313e544 1007/*
d62b8c6a 1008=for apidoc grok_atou
6313e544 1009
d62b8c6a 1010grok_atou is a safer replacement for atoi and strtol.
6313e544 1011
d62b8c6a
JH
1012grok_atou parses a C-style zero-byte terminated string, looking for
1013a decimal unsigned integer.
338aa8b0 1014
d62b8c6a
JH
1015Returns the unsigned integer, if a valid value can be parsed
1016from the beginning of the string.
f4379102 1017
d62b8c6a 1018Accepts only the decimal digits '0'..'9'.
6313e544 1019
d62b8c6a
JH
1020As opposed to atoi or strtol, grok_atou does NOT allow optional
1021leading whitespace, or negative inputs. If such features are
1022required, the calling code needs to explicitly implement those.
6313e544 1023
d62b8c6a 1024If a valid value cannot be parsed, returns either zero (if non-digits
75feedba 1025are met before any digits) or UV_MAX (if the value overflows).
6313e544 1026
d62b8c6a
JH
1027Note that extraneous leading zeros also count as an overflow
1028(meaning that only "0" is the zero).
338aa8b0 1029
d62b8c6a 1030On failure, the *endptr is also set to NULL, unless endptr is NULL.
338aa8b0
JH
1031
1032Trailing non-digit bytes are allowed if the endptr is non-NULL.
6313e544
JH
1033On return the *endptr will contain the pointer to the first non-digit byte.
1034
6313e544 1035If the endptr is NULL, the first non-digit byte MUST be
f4379102 1036the zero byte terminating the pv, or zero will be returned.
6313e544 1037
d62b8c6a
JH
1038Background: atoi has severe problems with illegal inputs, it cannot be
1039used for incremental parsing, and therefore should be avoided
1040atoi and strtol are also affected by locale settings, which can also be
1041seen as a bug (global state controlled by user environment).
1042
6313e544
JH
1043=cut
1044*/
1045
75feedba 1046UV
6313e544
JH
1047Perl_grok_atou(const char *pv, const char** endptr)
1048{
1049 const char* s = pv;
1050 const char** eptr;
1051 const char* end2; /* Used in case endptr is NULL. */
75feedba 1052 UV val = 0; /* The return value. */
6313e544
JH
1053
1054 PERL_ARGS_ASSERT_GROK_ATOU;
1055
1056 eptr = endptr ? endptr : &end2;
75feedba
JH
1057 if (isDIGIT(*s)) {
1058 /* Single-digit inputs are quite common. */
6313e544 1059 val = *s++ - '0';
75feedba
JH
1060 if (isDIGIT(*s)) {
1061 /* Extra leading zeros cause overflow. */
1062 if (val == 0) {
1063 *eptr = NULL;
1064 return UV_MAX;
1065 }
1066 while (isDIGIT(*s)) {
1067 /* This could be unrolled like in grok_number(), but
1068 * the expected uses of this are not speed-needy, and
1069 * unlikely to need full 64-bitness. */
1070 U8 digit = *s++ - '0';
945b524a
JH
1071 if (val < uv_max_div_10 ||
1072 (val == uv_max_div_10 && digit <= uv_max_mod_10)) {
75feedba
JH
1073 val = val * 10 + digit;
1074 } else {
6313e544 1075 *eptr = NULL;
75feedba 1076 return UV_MAX;
6313e544 1077 }
6313e544
JH
1078 }
1079 }
75feedba
JH
1080 }
1081 if (s == pv) {
1082 *eptr = NULL; /* If no progress, failed to parse anything. */
1083 return 0;
6313e544
JH
1084 }
1085 if (endptr == NULL && *s) {
1086 return 0; /* If endptr is NULL, no trailing non-digits allowed. */
1087 }
1088 *eptr = s;
1089 return val;
1090}
1091
a4eca1d4 1092#ifndef USE_QUADMATH
4801ca72 1093STATIC NV
98994639
HS
1094S_mulexp10(NV value, I32 exponent)
1095{
1096 NV result = 1.0;
1097 NV power = 10.0;
1098 bool negative = 0;
1099 I32 bit;
1100
1101 if (exponent == 0)
1102 return value;
659c4b96
DM
1103 if (value == 0)
1104 return (NV)0;
87032ba1 1105
24866caa 1106 /* On OpenVMS VAX we by default use the D_FLOAT double format,
67597c89 1107 * and that format does not have *easy* capabilities [1] for
24866caa
CB
1108 * overflowing doubles 'silently' as IEEE fp does. We also need
1109 * to support G_FLOAT on both VAX and Alpha, and though the exponent
1110 * range is much larger than D_FLOAT it still doesn't do silent
1111 * overflow. Therefore we need to detect early whether we would
1112 * overflow (this is the behaviour of the native string-to-float
1113 * conversion routines, and therefore of native applications, too).
67597c89 1114 *
24866caa
CB
1115 * [1] Trying to establish a condition handler to trap floating point
1116 * exceptions is not a good idea. */
87032ba1
JH
1117
1118 /* In UNICOS and in certain Cray models (such as T90) there is no
1119 * IEEE fp, and no way at all from C to catch fp overflows gracefully.
1120 * There is something you can do if you are willing to use some
1121 * inline assembler: the instruction is called DFI-- but that will
1122 * disable *all* floating point interrupts, a little bit too large
1123 * a hammer. Therefore we need to catch potential overflows before
1124 * it's too late. */
353813d9 1125
85bba25f 1126#if ((defined(VMS) && !defined(_IEEE_FP)) || defined(_UNICOS)) && defined(NV_MAX_10_EXP)
353813d9 1127 STMT_START {
c4420975 1128 const NV exp_v = log10(value);
353813d9
HS
1129 if (exponent >= NV_MAX_10_EXP || exponent + exp_v >= NV_MAX_10_EXP)
1130 return NV_MAX;
1131 if (exponent < 0) {
1132 if (-(exponent + exp_v) >= NV_MAX_10_EXP)
1133 return 0.0;
1134 while (-exponent >= NV_MAX_10_EXP) {
1135 /* combination does not overflow, but 10^(-exponent) does */
1136 value /= 10;
1137 ++exponent;
1138 }
1139 }
1140 } STMT_END;
87032ba1
JH
1141#endif
1142
353813d9
HS
1143 if (exponent < 0) {
1144 negative = 1;
1145 exponent = -exponent;
b27804d8
DM
1146#ifdef NV_MAX_10_EXP
1147 /* for something like 1234 x 10^-309, the action of calculating
1148 * the intermediate value 10^309 then returning 1234 / (10^309)
1149 * will fail, since 10^309 becomes infinity. In this case try to
1150 * refactor it as 123 / (10^308) etc.
1151 */
1152 while (value && exponent > NV_MAX_10_EXP) {
1153 exponent--;
1154 value /= 10;
1155 }
48853916
JH
1156 if (value == 0.0)
1157 return value;
b27804d8 1158#endif
353813d9 1159 }
c62e754c
JH
1160#if defined(__osf__)
1161 /* Even with cc -ieee + ieee_set_fp_control(IEEE_TRAP_ENABLE_INV)
1162 * Tru64 fp behavior on inf/nan is somewhat broken. Another way
1163 * to do this would be ieee_set_fp_control(IEEE_TRAP_ENABLE_OVF)
1164 * but that breaks another set of infnan.t tests. */
1165# define FP_OVERFLOWS_TO_ZERO
1166#endif
98994639
HS
1167 for (bit = 1; exponent; bit <<= 1) {
1168 if (exponent & bit) {
1169 exponent ^= bit;
1170 result *= power;
c62e754c
JH
1171#ifdef FP_OVERFLOWS_TO_ZERO
1172 if (result == 0)
1173 return value < 0 ? -NV_INF : NV_INF;
1174#endif
236f0012
CB
1175 /* Floating point exceptions are supposed to be turned off,
1176 * but if we're obviously done, don't risk another iteration.
1177 */
1178 if (exponent == 0) break;
98994639
HS
1179 }
1180 power *= power;
1181 }
1182 return negative ? value / result : value * result;
1183}
a4eca1d4 1184#endif /* #ifndef USE_QUADMATH */
98994639
HS
1185
1186NV
1187Perl_my_atof(pTHX_ const char* s)
1188{
1189 NV x = 0.0;
a4eca1d4
JH
1190#ifdef USE_QUADMATH
1191 Perl_my_atof2(aTHX_ s, &x);
1192 return x;
1193#else
1194# ifdef USE_LOCALE_NUMERIC
7918f24d
NC
1195 PERL_ARGS_ASSERT_MY_ATOF;
1196
a2287a13
KW
1197 {
1198 DECLARE_STORE_LC_NUMERIC_SET_TO_NEEDED();
d6ded950 1199 if (PL_numeric_radix_sv && IN_LC(LC_NUMERIC)) {
e4850248
KW
1200 const char *standard = NULL, *local = NULL;
1201 bool use_standard_radix;
98994639 1202
e4850248
KW
1203 /* Look through the string for the first thing that looks like a
1204 * decimal point: either the value in the current locale or the
1205 * standard fallback of '.'. The one which appears earliest in the
1206 * input string is the one that we should have atof look for. Note
1207 * that we have to determine this beforehand because on some
1208 * systems, Perl_atof2 is just a wrapper around the system's atof.
1209 * */
1210 standard = strchr(s, '.');
1211 local = strstr(s, SvPV_nolen(PL_numeric_radix_sv));
78787052 1212
e4850248 1213 use_standard_radix = standard && (!local || standard < local);
78787052 1214
e4850248
KW
1215 if (use_standard_radix)
1216 SET_NUMERIC_STANDARD();
78787052 1217
e4850248 1218 Perl_atof2(s, x);
78787052 1219
e4850248
KW
1220 if (use_standard_radix)
1221 SET_NUMERIC_LOCAL();
1222 }
1223 else
1224 Perl_atof2(s, x);
a2287a13
KW
1225 RESTORE_LC_NUMERIC();
1226 }
a4eca1d4 1227# else
a36244b7 1228 Perl_atof2(s, x);
a4eca1d4 1229# endif
98994639
HS
1230#endif
1231 return x;
1232}
1233
829757a4 1234static char*
5563f457 1235S_my_atof_infnan(pTHX_ const char* s, bool negative, const char* send, NV* value)
829757a4
JH
1236{
1237 const char *p0 = negative ? s - 1 : s;
1238 const char *p = p0;
1239 int infnan = grok_infnan(&p, send);
1240 if (infnan && p != p0) {
1241 /* If we can generate inf/nan directly, let's do so. */
1242#ifdef NV_INF
1243 if ((infnan & IS_NUMBER_INFINITY)) {
1244 *value = (infnan & IS_NUMBER_NEG) ? -NV_INF: NV_INF;
1245 return (char*)p;
1246 }
1247#endif
1248#ifdef NV_NAN
1249 if ((infnan & IS_NUMBER_NAN)) {
1250 *value = NV_NAN;
1251 return (char*)p;
1252 }
1253#endif
1254#ifdef Perl_strtod
68611e6f 1255 /* If still here, we didn't have either NV_INF or NV_NAN,
829757a4
JH
1256 * and can try falling back to native strtod/strtold.
1257 *
68611e6f
JH
1258 * (Though, are our NV_INF or NV_NAN ever not defined?)
1259 *
829757a4
JH
1260 * The native interface might not recognize all the possible
1261 * inf/nan strings Perl recognizes. What we can try
1262 * is to try faking the input. We will try inf/-inf/nan
1263 * as the most promising/portable input. */
1264 {
1265 const char* fake = NULL;
1266 char* endp;
1267 NV nv;
1268 if ((infnan & IS_NUMBER_INFINITY)) {
1269 fake = ((infnan & IS_NUMBER_NEG)) ? "-inf" : "inf";
1270 }
1271 else if ((infnan & IS_NUMBER_NAN)) {
1272 fake = "nan";
1273 }
1274 assert(fake);
1275 nv = Perl_strtod(fake, &endp);
1276 if (fake != endp) {
1277 if ((infnan & IS_NUMBER_INFINITY)) {
1278#ifdef Perl_isinf
1279 if (Perl_isinf(nv))
1280 *value = nv;
1281#else
1282 /* last resort, may generate SIGFPE */
1283 *value = Perl_exp((NV)1e9);
1284 if ((infnan & IS_NUMBER_NEG))
1285 *value = -*value;
1286#endif
1287 return (char*)p; /* p, not endp */
1288 }
1289 else if ((infnan & IS_NUMBER_NAN)) {
1290#ifdef Perl_isnan
1291 if (Perl_isnan(nv))
1292 *value = nv;
1293#else
1294 /* last resort, may generate SIGFPE */
1295 *value = Perl_log((NV)-1.0);
1296#endif
1297 return (char*)p; /* p, not endp */
1298 }
1299 }
1300 }
1301#endif /* #ifdef Perl_strtod */
1302 }
1303 return NULL;
1304}
1305
98994639
HS
1306char*
1307Perl_my_atof2(pTHX_ const char* orig, NV* value)
1308{
e1ec3a88 1309 const char* s = orig;
a4eca1d4
JH
1310 NV result[3] = {0.0, 0.0, 0.0};
1311#if defined(USE_PERL_ATOF) || defined(USE_QUADMATH)
ae776a2c 1312 const char* send = s + strlen(orig); /* one past the last */
a4eca1d4
JH
1313 bool negative = 0;
1314#endif
1315#if defined(USE_PERL_ATOF) && !defined(USE_QUADMATH)
1316 UV accumulator[2] = {0,0}; /* before/after dp */
8194bf88 1317 bool seen_digit = 0;
20f6aaab
AS
1318 I32 exp_adjust[2] = {0,0};
1319 I32 exp_acc[2] = {-1, -1};
1320 /* the current exponent adjust for the accumulators */
98994639 1321 I32 exponent = 0;
8194bf88 1322 I32 seen_dp = 0;
20f6aaab
AS
1323 I32 digit = 0;
1324 I32 old_digit = 0;
8194bf88 1325 I32 sig_digits = 0; /* noof significant digits seen so far */
a4eca1d4 1326#endif
8194bf88 1327
a4eca1d4 1328#if defined(USE_PERL_ATOF) || defined(USE_QUADMATH)
7918f24d
NC
1329 PERL_ARGS_ASSERT_MY_ATOF2;
1330
a4eca1d4
JH
1331 /* leading whitespace */
1332 while (isSPACE(*s))
1333 ++s;
1334
1335 /* sign */
1336 switch (*s) {
1337 case '-':
1338 negative = 1;
1339 /* FALLTHROUGH */
1340 case '+':
1341 ++s;
1342 }
1343#endif
1344
1345#ifdef USE_QUADMATH
1346 {
1347 char* endp;
1348 if ((endp = S_my_atof_infnan(s, negative, send, value)))
1349 return endp;
1350 result[2] = strtoflt128(s, &endp);
1351 if (s != endp) {
1352 *value = negative ? -result[2] : result[2];
1353 return endp;
1354 }
1355 return NULL;
1356 }
1357#elif defined(USE_PERL_ATOF)
1358
8194bf88
DM
1359/* There is no point in processing more significant digits
1360 * than the NV can hold. Note that NV_DIG is a lower-bound value,
1361 * while we need an upper-bound value. We add 2 to account for this;
1362 * since it will have been conservative on both the first and last digit.
1363 * For example a 32-bit mantissa with an exponent of 4 would have
1364 * exact values in the set
1365 * 4
1366 * 8
1367 * ..
1368 * 17179869172
1369 * 17179869176
1370 * 17179869180
1371 *
1372 * where for the purposes of calculating NV_DIG we would have to discount
1373 * both the first and last digit, since neither can hold all values from
1374 * 0..9; but for calculating the value we must examine those two digits.
1375 */
ffa277e5
AS
1376#ifdef MAX_SIG_DIG_PLUS
1377 /* It is not necessarily the case that adding 2 to NV_DIG gets all the
1378 possible digits in a NV, especially if NVs are not IEEE compliant
1379 (e.g., long doubles on IRIX) - Allen <allens@cpan.org> */
1380# define MAX_SIG_DIGITS (NV_DIG+MAX_SIG_DIG_PLUS)
1381#else
1382# define MAX_SIG_DIGITS (NV_DIG+2)
1383#endif
8194bf88
DM
1384
1385/* the max number we can accumulate in a UV, and still safely do 10*N+9 */
1386#define MAX_ACCUMULATE ( (UV) ((UV_MAX - 9)/10))
98994639 1387
ae776a2c 1388 {
829757a4 1389 const char* endp;
5563f457 1390 if ((endp = S_my_atof_infnan(aTHX_ s, negative, send, value)))
829757a4 1391 return (char*)endp;
ae776a2c 1392 }
2b54f59f 1393
8194bf88
DM
1394 /* we accumulate digits into an integer; when this becomes too
1395 * large, we add the total to NV and start again */
98994639 1396
8194bf88
DM
1397 while (1) {
1398 if (isDIGIT(*s)) {
1399 seen_digit = 1;
20f6aaab 1400 old_digit = digit;
8194bf88 1401 digit = *s++ - '0';
20f6aaab
AS
1402 if (seen_dp)
1403 exp_adjust[1]++;
98994639 1404
8194bf88
DM
1405 /* don't start counting until we see the first significant
1406 * digit, eg the 5 in 0.00005... */
1407 if (!sig_digits && digit == 0)
1408 continue;
1409
1410 if (++sig_digits > MAX_SIG_DIGITS) {
98994639 1411 /* limits of precision reached */
20f6aaab
AS
1412 if (digit > 5) {
1413 ++accumulator[seen_dp];
1414 } else if (digit == 5) {
1415 if (old_digit % 2) { /* round to even - Allen */
1416 ++accumulator[seen_dp];
1417 }
1418 }
1419 if (seen_dp) {
1420 exp_adjust[1]--;
1421 } else {
1422 exp_adjust[0]++;
1423 }
8194bf88 1424 /* skip remaining digits */
98994639 1425 while (isDIGIT(*s)) {
98994639 1426 ++s;
20f6aaab
AS
1427 if (! seen_dp) {
1428 exp_adjust[0]++;
1429 }
98994639
HS
1430 }
1431 /* warn of loss of precision? */
98994639 1432 }
8194bf88 1433 else {
20f6aaab 1434 if (accumulator[seen_dp] > MAX_ACCUMULATE) {
8194bf88 1435 /* add accumulator to result and start again */
20f6aaab
AS
1436 result[seen_dp] = S_mulexp10(result[seen_dp],
1437 exp_acc[seen_dp])
1438 + (NV)accumulator[seen_dp];
1439 accumulator[seen_dp] = 0;
1440 exp_acc[seen_dp] = 0;
98994639 1441 }
20f6aaab
AS
1442 accumulator[seen_dp] = accumulator[seen_dp] * 10 + digit;
1443 ++exp_acc[seen_dp];
98994639 1444 }
8194bf88 1445 }
e1ec3a88 1446 else if (!seen_dp && GROK_NUMERIC_RADIX(&s, send)) {
8194bf88 1447 seen_dp = 1;
20f6aaab 1448 if (sig_digits > MAX_SIG_DIGITS) {
c86f7df5 1449 do {
20f6aaab 1450 ++s;
c86f7df5 1451 } while (isDIGIT(*s));
20f6aaab
AS
1452 break;
1453 }
8194bf88
DM
1454 }
1455 else {
1456 break;
98994639
HS
1457 }
1458 }
1459
20f6aaab
AS
1460 result[0] = S_mulexp10(result[0], exp_acc[0]) + (NV)accumulator[0];
1461 if (seen_dp) {
1462 result[1] = S_mulexp10(result[1], exp_acc[1]) + (NV)accumulator[1];
1463 }
98994639 1464
305b8651 1465 if (seen_digit && (isALPHA_FOLD_EQ(*s, 'e'))) {
98994639
HS
1466 bool expnegative = 0;
1467
1468 ++s;
1469 switch (*s) {
1470 case '-':
1471 expnegative = 1;
924ba076 1472 /* FALLTHROUGH */
98994639
HS
1473 case '+':
1474 ++s;
1475 }
1476 while (isDIGIT(*s))
1477 exponent = exponent * 10 + (*s++ - '0');
1478 if (expnegative)
1479 exponent = -exponent;
1480 }
1481
20f6aaab
AS
1482
1483
98994639 1484 /* now apply the exponent */
20f6aaab
AS
1485
1486 if (seen_dp) {
1487 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0])
1488 + S_mulexp10(result[1],exponent-exp_adjust[1]);
1489 } else {
1490 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0]);
1491 }
98994639
HS
1492
1493 /* now apply the sign */
1494 if (negative)
20f6aaab 1495 result[2] = -result[2];
a36244b7 1496#endif /* USE_PERL_ATOF */
20f6aaab 1497 *value = result[2];
73d840c0 1498 return (char *)s;
98994639
HS
1499}
1500
5d34af89 1501/*
3d9d9213 1502=for apidoc isinfnan
5d34af89
JH
1503
1504Perl_isinfnan() is utility function that returns true if the NV
1505argument is either an infinity or a NaN, false otherwise. To test
1506in more detail, use Perl_isinf() and Perl_isnan().
1507
68611e6f
JH
1508This is also the logical inverse of Perl_isfinite().
1509
5d34af89
JH
1510=cut
1511*/
1cd88304
JH
1512bool
1513Perl_isinfnan(NV nv)
1514{
1515#ifdef Perl_isinf
1516 if (Perl_isinf(nv))
1517 return TRUE;
1518#endif
1519#ifdef Perl_isnan
1520 if (Perl_isnan(nv))
1521 return TRUE;
1522#endif
1523 return FALSE;
1524}
1525
354b74ae
FC
1526/*
1527=for apidoc
1528
1529Checks whether the argument would be either an infinity or NaN when used
1530as a number, but is careful not to trigger non-numeric or uninitialized
1531warnings. it assumes the caller has done SvGETMAGIC(sv) already.
1532
1533=cut
1534*/
1535
1536bool
1537Perl_isinfnansv(pTHX_ SV *sv)
1538{
1539 PERL_ARGS_ASSERT_ISINFNANSV;
1540 if (!SvOK(sv))
1541 return FALSE;
1542 if (SvNOKp(sv))
1543 return Perl_isinfnan(SvNVX(sv));
1544 if (SvIOKp(sv))
1545 return FALSE;
1546 {
1547 STRLEN len;
1548 const char *s = SvPV_nomg_const(sv, len);
1549 return cBOOL(grok_infnan(&s, s+len));
1550 }
1551}
1552
d67dac15 1553#ifndef HAS_MODFL
68611e6f
JH
1554/* C99 has truncl, pre-C99 Solaris had aintl. We can use either with
1555 * copysignl to emulate modfl, which is in some platforms missing or
1556 * broken. */
d67dac15
JH
1557# if defined(HAS_TRUNCL) && defined(HAS_COPYSIGNL)
1558long double
1559Perl_my_modfl(long double x, long double *ip)
1560{
68611e6f
JH
1561 *ip = truncl(x);
1562 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
d67dac15
JH
1563}
1564# elif defined(HAS_AINTL) && defined(HAS_COPYSIGNL)
55954f19
JH
1565long double
1566Perl_my_modfl(long double x, long double *ip)
1567{
68611e6f
JH
1568 *ip = aintl(x);
1569 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
55954f19 1570}
d67dac15 1571# endif
55954f19
JH
1572#endif
1573
7b9b7dff 1574/* Similarly, with ilogbl and scalbnl we can emulate frexpl. */
55954f19
JH
1575#if ! defined(HAS_FREXPL) && defined(HAS_ILOGBL) && defined(HAS_SCALBNL)
1576long double
1577Perl_my_frexpl(long double x, int *e) {
68611e6f
JH
1578 *e = x == 0.0L ? 0 : ilogbl(x) + 1;
1579 return (scalbnl(x, -*e));
55954f19
JH
1580}
1581#endif
66610fdd
RGS
1582
1583/*
ed140128
AD
1584=for apidoc Perl_signbit
1585
1586Return a non-zero integer if the sign bit on an NV is set, and 0 if
1587it is not.
1588
1589If Configure detects this system has a signbit() that will work with
1590our NVs, then we just use it via the #define in perl.h. Otherwise,
8b7fad81
JH
1591fall back on this implementation. The main use of this function
1592is catching -0.0.
ed140128
AD
1593
1594Configure notes: This function is called 'Perl_signbit' instead of a
1595plain 'signbit' because it is easy to imagine a system having a signbit()
1596function or macro that doesn't happen to work with our particular choice
1597of NVs. We shouldn't just re-#define signbit as Perl_signbit and expect
1598the standard system headers to be happy. Also, this is a no-context
1599function (no pTHX_) because Perl_signbit() is usually re-#defined in
1600perl.h as a simple macro call to the system's signbit().
1601Users should just always call Perl_signbit().
1602
1603=cut
1604*/
1605#if !defined(HAS_SIGNBIT)
1606int
1607Perl_signbit(NV x) {
8b7fad81
JH
1608# ifdef Perl_fp_class_nzero
1609 if (x == 0)
1610 return Perl_fp_class_nzero(x);
8b7fad81 1611# endif
3585840c 1612 return (x < 0.0) ? 1 : 0;
ed140128
AD
1613}
1614#endif
1615
1616/*
66610fdd
RGS
1617 * Local variables:
1618 * c-indentation-style: bsd
1619 * c-basic-offset: 4
14d04a33 1620 * indent-tabs-mode: nil
66610fdd
RGS
1621 * End:
1622 *
14d04a33 1623 * ex: set ts=8 sts=4 sw=4 et:
37442d52 1624 */