This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
Update IO-Compress to CPAN version 2.066
[perl5.git] / numeric.c
CommitLineData
98994639
HS
1/* numeric.c
2 *
663f364b 3 * Copyright (C) 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001,
1129b882 4 * 2002, 2003, 2004, 2005, 2006, 2007, 2008 by Larry Wall and others
98994639
HS
5 *
6 * You may distribute under the terms of either the GNU General Public
7 * License or the Artistic License, as specified in the README file.
8 *
9 */
10
11/*
4ac71550
TC
12 * "That only makes eleven (plus one mislaid) and not fourteen,
13 * unless wizards count differently to other people." --Beorn
14 *
15 * [p.115 of _The Hobbit_: "Queer Lodgings"]
98994639
HS
16 */
17
ccfc67b7
JH
18/*
19=head1 Numeric functions
166f8a29 20
7fefc6c1
KW
21=cut
22
166f8a29
DM
23This file contains all the stuff needed by perl for manipulating numeric
24values, including such things as replacements for the OS's atof() function
25
ccfc67b7
JH
26*/
27
98994639
HS
28#include "EXTERN.h"
29#define PERL_IN_NUMERIC_C
30#include "perl.h"
31
32U32
ddeaf645 33Perl_cast_ulong(NV f)
98994639
HS
34{
35 if (f < 0.0)
36 return f < I32_MIN ? (U32) I32_MIN : (U32)(I32) f;
37 if (f < U32_MAX_P1) {
38#if CASTFLAGS & 2
39 if (f < U32_MAX_P1_HALF)
40 return (U32) f;
41 f -= U32_MAX_P1_HALF;
42 return ((U32) f) | (1 + U32_MAX >> 1);
43#else
44 return (U32) f;
45#endif
46 }
47 return f > 0 ? U32_MAX : 0 /* NaN */;
48}
49
50I32
ddeaf645 51Perl_cast_i32(NV f)
98994639
HS
52{
53 if (f < I32_MAX_P1)
54 return f < I32_MIN ? I32_MIN : (I32) f;
55 if (f < U32_MAX_P1) {
56#if CASTFLAGS & 2
57 if (f < U32_MAX_P1_HALF)
58 return (I32)(U32) f;
59 f -= U32_MAX_P1_HALF;
60 return (I32)(((U32) f) | (1 + U32_MAX >> 1));
61#else
62 return (I32)(U32) f;
63#endif
64 }
65 return f > 0 ? (I32)U32_MAX : 0 /* NaN */;
66}
67
68IV
ddeaf645 69Perl_cast_iv(NV f)
98994639
HS
70{
71 if (f < IV_MAX_P1)
72 return f < IV_MIN ? IV_MIN : (IV) f;
73 if (f < UV_MAX_P1) {
74#if CASTFLAGS & 2
75 /* For future flexibility allowing for sizeof(UV) >= sizeof(IV) */
76 if (f < UV_MAX_P1_HALF)
77 return (IV)(UV) f;
78 f -= UV_MAX_P1_HALF;
79 return (IV)(((UV) f) | (1 + UV_MAX >> 1));
80#else
81 return (IV)(UV) f;
82#endif
83 }
84 return f > 0 ? (IV)UV_MAX : 0 /* NaN */;
85}
86
87UV
ddeaf645 88Perl_cast_uv(NV f)
98994639
HS
89{
90 if (f < 0.0)
91 return f < IV_MIN ? (UV) IV_MIN : (UV)(IV) f;
92 if (f < UV_MAX_P1) {
93#if CASTFLAGS & 2
94 if (f < UV_MAX_P1_HALF)
95 return (UV) f;
96 f -= UV_MAX_P1_HALF;
97 return ((UV) f) | (1 + UV_MAX >> 1);
98#else
99 return (UV) f;
100#endif
101 }
102 return f > 0 ? UV_MAX : 0 /* NaN */;
103}
104
53305cf1
NC
105/*
106=for apidoc grok_bin
98994639 107
53305cf1
NC
108converts a string representing a binary number to numeric form.
109
110On entry I<start> and I<*len> give the string to scan, I<*flags> gives
111conversion flags, and I<result> should be NULL or a pointer to an NV.
112The scan stops at the end of the string, or the first invalid character.
7b667b5f
MHM
113Unless C<PERL_SCAN_SILENT_ILLDIGIT> is set in I<*flags>, encountering an
114invalid character will also trigger a warning.
115On return I<*len> is set to the length of the scanned string,
116and I<*flags> gives output flags.
53305cf1 117
7fc63493 118If the value is <= C<UV_MAX> it is returned as a UV, the output flags are clear,
72d33970 119and nothing is written to I<*result>. If the value is > UV_MAX C<grok_bin>
53305cf1
NC
120returns UV_MAX, sets C<PERL_SCAN_GREATER_THAN_UV_MAX> in the output flags,
121and writes the value to I<*result> (or the value is discarded if I<result>
122is NULL).
123
7b667b5f 124The binary number may optionally be prefixed with "0b" or "b" unless
72d33970 125C<PERL_SCAN_DISALLOW_PREFIX> is set in I<*flags> on entry. If
a4c04bdc 126C<PERL_SCAN_ALLOW_UNDERSCORES> is set in I<*flags> then the binary
53305cf1
NC
127number may use '_' characters to separate digits.
128
129=cut
02470786
KW
130
131Not documented yet because experimental is C<PERL_SCAN_SILENT_NON_PORTABLE
132which suppresses any message for non-portable numbers that are still valid
133on this platform.
53305cf1
NC
134 */
135
136UV
7918f24d
NC
137Perl_grok_bin(pTHX_ const char *start, STRLEN *len_p, I32 *flags, NV *result)
138{
53305cf1
NC
139 const char *s = start;
140 STRLEN len = *len_p;
141 UV value = 0;
142 NV value_nv = 0;
143
144 const UV max_div_2 = UV_MAX / 2;
f2338a2e 145 const bool allow_underscores = cBOOL(*flags & PERL_SCAN_ALLOW_UNDERSCORES);
53305cf1 146 bool overflowed = FALSE;
7fc63493 147 char bit;
53305cf1 148
7918f24d
NC
149 PERL_ARGS_ASSERT_GROK_BIN;
150
a4c04bdc
NC
151 if (!(*flags & PERL_SCAN_DISALLOW_PREFIX)) {
152 /* strip off leading b or 0b.
153 for compatibility silently suffer "b" and "0b" as valid binary
154 numbers. */
155 if (len >= 1) {
305b8651 156 if (isALPHA_FOLD_EQ(s[0], 'b')) {
a4c04bdc
NC
157 s++;
158 len--;
159 }
305b8651 160 else if (len >= 2 && s[0] == '0' && (isALPHA_FOLD_EQ(s[1], 'b'))) {
a4c04bdc
NC
161 s+=2;
162 len-=2;
163 }
164 }
53305cf1
NC
165 }
166
7fc63493 167 for (; len-- && (bit = *s); s++) {
53305cf1
NC
168 if (bit == '0' || bit == '1') {
169 /* Write it in this wonky order with a goto to attempt to get the
170 compiler to make the common case integer-only loop pretty tight.
171 With gcc seems to be much straighter code than old scan_bin. */
172 redo:
173 if (!overflowed) {
174 if (value <= max_div_2) {
175 value = (value << 1) | (bit - '0');
176 continue;
177 }
178 /* Bah. We're just overflowed. */
dcbac5bb 179 /* diag_listed_as: Integer overflow in %s number */
9b387841
NC
180 Perl_ck_warner_d(aTHX_ packWARN(WARN_OVERFLOW),
181 "Integer overflow in binary number");
53305cf1
NC
182 overflowed = TRUE;
183 value_nv = (NV) value;
184 }
185 value_nv *= 2.0;
98994639 186 /* If an NV has not enough bits in its mantissa to
d1be9408 187 * represent a UV this summing of small low-order numbers
98994639
HS
188 * is a waste of time (because the NV cannot preserve
189 * the low-order bits anyway): we could just remember when
53305cf1 190 * did we overflow and in the end just multiply value_nv by the
98994639 191 * right amount. */
53305cf1
NC
192 value_nv += (NV)(bit - '0');
193 continue;
194 }
195 if (bit == '_' && len && allow_underscores && (bit = s[1])
196 && (bit == '0' || bit == '1'))
98994639
HS
197 {
198 --len;
199 ++s;
53305cf1 200 goto redo;
98994639 201 }
a2a5de95
NC
202 if (!(*flags & PERL_SCAN_SILENT_ILLDIGIT))
203 Perl_ck_warner(aTHX_ packWARN(WARN_DIGIT),
204 "Illegal binary digit '%c' ignored", *s);
53305cf1 205 break;
98994639 206 }
53305cf1
NC
207
208 if ( ( overflowed && value_nv > 4294967295.0)
98994639 209#if UVSIZE > 4
02470786
KW
210 || (!overflowed && value > 0xffffffff
211 && ! (*flags & PERL_SCAN_SILENT_NON_PORTABLE))
98994639
HS
212#endif
213 ) {
a2a5de95
NC
214 Perl_ck_warner(aTHX_ packWARN(WARN_PORTABLE),
215 "Binary number > 0b11111111111111111111111111111111 non-portable");
53305cf1
NC
216 }
217 *len_p = s - start;
218 if (!overflowed) {
219 *flags = 0;
220 return value;
98994639 221 }
53305cf1
NC
222 *flags = PERL_SCAN_GREATER_THAN_UV_MAX;
223 if (result)
224 *result = value_nv;
225 return UV_MAX;
98994639
HS
226}
227
53305cf1
NC
228/*
229=for apidoc grok_hex
230
231converts a string representing a hex number to numeric form.
232
c2da02fc 233On entry I<start> and I<*len_p> give the string to scan, I<*flags> gives
53305cf1 234conversion flags, and I<result> should be NULL or a pointer to an NV.
7b667b5f
MHM
235The scan stops at the end of the string, or the first invalid character.
236Unless C<PERL_SCAN_SILENT_ILLDIGIT> is set in I<*flags>, encountering an
237invalid character will also trigger a warning.
238On return I<*len> is set to the length of the scanned string,
239and I<*flags> gives output flags.
53305cf1
NC
240
241If the value is <= UV_MAX it is returned as a UV, the output flags are clear,
72d33970 242and nothing is written to I<*result>. If the value is > UV_MAX C<grok_hex>
53305cf1
NC
243returns UV_MAX, sets C<PERL_SCAN_GREATER_THAN_UV_MAX> in the output flags,
244and writes the value to I<*result> (or the value is discarded if I<result>
245is NULL).
246
d1be9408 247The hex number may optionally be prefixed with "0x" or "x" unless
72d33970 248C<PERL_SCAN_DISALLOW_PREFIX> is set in I<*flags> on entry. If
a4c04bdc 249C<PERL_SCAN_ALLOW_UNDERSCORES> is set in I<*flags> then the hex
53305cf1
NC
250number may use '_' characters to separate digits.
251
252=cut
02470786
KW
253
254Not documented yet because experimental is C<PERL_SCAN_SILENT_NON_PORTABLE
baf48926 255which suppresses any message for non-portable numbers, but which are valid
02470786 256on this platform.
53305cf1
NC
257 */
258
259UV
7918f24d
NC
260Perl_grok_hex(pTHX_ const char *start, STRLEN *len_p, I32 *flags, NV *result)
261{
53305cf1
NC
262 const char *s = start;
263 STRLEN len = *len_p;
264 UV value = 0;
265 NV value_nv = 0;
53305cf1 266 const UV max_div_16 = UV_MAX / 16;
f2338a2e 267 const bool allow_underscores = cBOOL(*flags & PERL_SCAN_ALLOW_UNDERSCORES);
53305cf1 268 bool overflowed = FALSE;
98994639 269
7918f24d
NC
270 PERL_ARGS_ASSERT_GROK_HEX;
271
a4c04bdc
NC
272 if (!(*flags & PERL_SCAN_DISALLOW_PREFIX)) {
273 /* strip off leading x or 0x.
274 for compatibility silently suffer "x" and "0x" as valid hex numbers.
275 */
276 if (len >= 1) {
305b8651 277 if (isALPHA_FOLD_EQ(s[0], 'x')) {
a4c04bdc
NC
278 s++;
279 len--;
280 }
305b8651 281 else if (len >= 2 && s[0] == '0' && (isALPHA_FOLD_EQ(s[1], 'x'))) {
a4c04bdc
NC
282 s+=2;
283 len-=2;
284 }
285 }
98994639
HS
286 }
287
288 for (; len-- && *s; s++) {
626ef089 289 if (isXDIGIT(*s)) {
53305cf1
NC
290 /* Write it in this wonky order with a goto to attempt to get the
291 compiler to make the common case integer-only loop pretty tight.
292 With gcc seems to be much straighter code than old scan_hex. */
293 redo:
294 if (!overflowed) {
295 if (value <= max_div_16) {
626ef089 296 value = (value << 4) | XDIGIT_VALUE(*s);
53305cf1
NC
297 continue;
298 }
299 /* Bah. We're just overflowed. */
dcbac5bb 300 /* diag_listed_as: Integer overflow in %s number */
9b387841
NC
301 Perl_ck_warner_d(aTHX_ packWARN(WARN_OVERFLOW),
302 "Integer overflow in hexadecimal number");
53305cf1
NC
303 overflowed = TRUE;
304 value_nv = (NV) value;
305 }
306 value_nv *= 16.0;
307 /* If an NV has not enough bits in its mantissa to
d1be9408 308 * represent a UV this summing of small low-order numbers
53305cf1
NC
309 * is a waste of time (because the NV cannot preserve
310 * the low-order bits anyway): we could just remember when
311 * did we overflow and in the end just multiply value_nv by the
312 * right amount of 16-tuples. */
626ef089 313 value_nv += (NV) XDIGIT_VALUE(*s);
53305cf1
NC
314 continue;
315 }
316 if (*s == '_' && len && allow_underscores && s[1]
626ef089 317 && isXDIGIT(s[1]))
98994639
HS
318 {
319 --len;
320 ++s;
53305cf1 321 goto redo;
98994639 322 }
a2a5de95
NC
323 if (!(*flags & PERL_SCAN_SILENT_ILLDIGIT))
324 Perl_ck_warner(aTHX_ packWARN(WARN_DIGIT),
53305cf1
NC
325 "Illegal hexadecimal digit '%c' ignored", *s);
326 break;
327 }
328
329 if ( ( overflowed && value_nv > 4294967295.0)
330#if UVSIZE > 4
02470786
KW
331 || (!overflowed && value > 0xffffffff
332 && ! (*flags & PERL_SCAN_SILENT_NON_PORTABLE))
53305cf1
NC
333#endif
334 ) {
a2a5de95
NC
335 Perl_ck_warner(aTHX_ packWARN(WARN_PORTABLE),
336 "Hexadecimal number > 0xffffffff non-portable");
53305cf1
NC
337 }
338 *len_p = s - start;
339 if (!overflowed) {
340 *flags = 0;
341 return value;
342 }
343 *flags = PERL_SCAN_GREATER_THAN_UV_MAX;
344 if (result)
345 *result = value_nv;
346 return UV_MAX;
347}
348
349/*
350=for apidoc grok_oct
351
7b667b5f
MHM
352converts a string representing an octal number to numeric form.
353
354On entry I<start> and I<*len> give the string to scan, I<*flags> gives
355conversion flags, and I<result> should be NULL or a pointer to an NV.
356The scan stops at the end of the string, or the first invalid character.
357Unless C<PERL_SCAN_SILENT_ILLDIGIT> is set in I<*flags>, encountering an
154bd527 3588 or 9 will also trigger a warning.
7b667b5f
MHM
359On return I<*len> is set to the length of the scanned string,
360and I<*flags> gives output flags.
361
362If the value is <= UV_MAX it is returned as a UV, the output flags are clear,
72d33970 363and nothing is written to I<*result>. If the value is > UV_MAX C<grok_oct>
7b667b5f
MHM
364returns UV_MAX, sets C<PERL_SCAN_GREATER_THAN_UV_MAX> in the output flags,
365and writes the value to I<*result> (or the value is discarded if I<result>
366is NULL).
367
368If C<PERL_SCAN_ALLOW_UNDERSCORES> is set in I<*flags> then the octal
369number may use '_' characters to separate digits.
53305cf1
NC
370
371=cut
02470786 372
333ae27c
KW
373Not documented yet because experimental is C<PERL_SCAN_SILENT_NON_PORTABLE>
374which suppresses any message for non-portable numbers, but which are valid
02470786 375on this platform.
53305cf1
NC
376 */
377
378UV
7918f24d
NC
379Perl_grok_oct(pTHX_ const char *start, STRLEN *len_p, I32 *flags, NV *result)
380{
53305cf1
NC
381 const char *s = start;
382 STRLEN len = *len_p;
383 UV value = 0;
384 NV value_nv = 0;
53305cf1 385 const UV max_div_8 = UV_MAX / 8;
f2338a2e 386 const bool allow_underscores = cBOOL(*flags & PERL_SCAN_ALLOW_UNDERSCORES);
53305cf1
NC
387 bool overflowed = FALSE;
388
7918f24d
NC
389 PERL_ARGS_ASSERT_GROK_OCT;
390
53305cf1 391 for (; len-- && *s; s++) {
626ef089 392 if (isOCTAL(*s)) {
53305cf1
NC
393 /* Write it in this wonky order with a goto to attempt to get the
394 compiler to make the common case integer-only loop pretty tight.
395 */
396 redo:
397 if (!overflowed) {
398 if (value <= max_div_8) {
626ef089 399 value = (value << 3) | OCTAL_VALUE(*s);
53305cf1
NC
400 continue;
401 }
402 /* Bah. We're just overflowed. */
dcbac5bb 403 /* diag_listed_as: Integer overflow in %s number */
9b387841
NC
404 Perl_ck_warner_d(aTHX_ packWARN(WARN_OVERFLOW),
405 "Integer overflow in octal number");
53305cf1
NC
406 overflowed = TRUE;
407 value_nv = (NV) value;
408 }
409 value_nv *= 8.0;
98994639 410 /* If an NV has not enough bits in its mantissa to
d1be9408 411 * represent a UV this summing of small low-order numbers
98994639
HS
412 * is a waste of time (because the NV cannot preserve
413 * the low-order bits anyway): we could just remember when
53305cf1
NC
414 * did we overflow and in the end just multiply value_nv by the
415 * right amount of 8-tuples. */
626ef089 416 value_nv += (NV) OCTAL_VALUE(*s);
53305cf1
NC
417 continue;
418 }
626ef089
KW
419 if (*s == '_' && len && allow_underscores && isOCTAL(s[1])) {
420 --len;
421 ++s;
422 goto redo;
423 }
53305cf1 424 /* Allow \octal to work the DWIM way (that is, stop scanning
7b667b5f 425 * as soon as non-octal characters are seen, complain only if
626ef089
KW
426 * someone seems to want to use the digits eight and nine. Since we
427 * know it is not octal, then if isDIGIT, must be an 8 or 9). */
428 if (isDIGIT(*s)) {
a2a5de95
NC
429 if (!(*flags & PERL_SCAN_SILENT_ILLDIGIT))
430 Perl_ck_warner(aTHX_ packWARN(WARN_DIGIT),
431 "Illegal octal digit '%c' ignored", *s);
53305cf1
NC
432 }
433 break;
98994639 434 }
53305cf1
NC
435
436 if ( ( overflowed && value_nv > 4294967295.0)
98994639 437#if UVSIZE > 4
02470786
KW
438 || (!overflowed && value > 0xffffffff
439 && ! (*flags & PERL_SCAN_SILENT_NON_PORTABLE))
98994639
HS
440#endif
441 ) {
a2a5de95
NC
442 Perl_ck_warner(aTHX_ packWARN(WARN_PORTABLE),
443 "Octal number > 037777777777 non-portable");
53305cf1
NC
444 }
445 *len_p = s - start;
446 if (!overflowed) {
447 *flags = 0;
448 return value;
98994639 449 }
53305cf1
NC
450 *flags = PERL_SCAN_GREATER_THAN_UV_MAX;
451 if (result)
452 *result = value_nv;
453 return UV_MAX;
454}
455
456/*
457=for apidoc scan_bin
458
72d33970 459For backwards compatibility. Use C<grok_bin> instead.
53305cf1
NC
460
461=for apidoc scan_hex
462
72d33970 463For backwards compatibility. Use C<grok_hex> instead.
53305cf1
NC
464
465=for apidoc scan_oct
466
72d33970 467For backwards compatibility. Use C<grok_oct> instead.
53305cf1
NC
468
469=cut
470 */
471
472NV
73d840c0 473Perl_scan_bin(pTHX_ const char *start, STRLEN len, STRLEN *retlen)
53305cf1
NC
474{
475 NV rnv;
476 I32 flags = *retlen ? PERL_SCAN_ALLOW_UNDERSCORES : 0;
73d840c0 477 const UV ruv = grok_bin (start, &len, &flags, &rnv);
53305cf1 478
7918f24d
NC
479 PERL_ARGS_ASSERT_SCAN_BIN;
480
53305cf1
NC
481 *retlen = len;
482 return (flags & PERL_SCAN_GREATER_THAN_UV_MAX) ? rnv : (NV)ruv;
483}
484
485NV
73d840c0 486Perl_scan_oct(pTHX_ const char *start, STRLEN len, STRLEN *retlen)
53305cf1
NC
487{
488 NV rnv;
489 I32 flags = *retlen ? PERL_SCAN_ALLOW_UNDERSCORES : 0;
73d840c0 490 const UV ruv = grok_oct (start, &len, &flags, &rnv);
53305cf1 491
7918f24d
NC
492 PERL_ARGS_ASSERT_SCAN_OCT;
493
53305cf1
NC
494 *retlen = len;
495 return (flags & PERL_SCAN_GREATER_THAN_UV_MAX) ? rnv : (NV)ruv;
496}
497
498NV
73d840c0 499Perl_scan_hex(pTHX_ const char *start, STRLEN len, STRLEN *retlen)
53305cf1
NC
500{
501 NV rnv;
502 I32 flags = *retlen ? PERL_SCAN_ALLOW_UNDERSCORES : 0;
73d840c0 503 const UV ruv = grok_hex (start, &len, &flags, &rnv);
53305cf1 504
7918f24d
NC
505 PERL_ARGS_ASSERT_SCAN_HEX;
506
53305cf1
NC
507 *retlen = len;
508 return (flags & PERL_SCAN_GREATER_THAN_UV_MAX) ? rnv : (NV)ruv;
98994639
HS
509}
510
511/*
512=for apidoc grok_numeric_radix
513
514Scan and skip for a numeric decimal separator (radix).
515
516=cut
517 */
518bool
519Perl_grok_numeric_radix(pTHX_ const char **sp, const char *send)
520{
521#ifdef USE_LOCALE_NUMERIC
7918f24d
NC
522 PERL_ARGS_ASSERT_GROK_NUMERIC_RADIX;
523
d6ded950 524 if (IN_LC(LC_NUMERIC)) {
21431899
KW
525 DECLARE_STORE_LC_NUMERIC_SET_TO_NEEDED();
526 if (PL_numeric_radix_sv) {
c5a7e38e
KW
527 STRLEN len;
528 const char * const radix = SvPV(PL_numeric_radix_sv, len);
529 if (*sp + len <= send && memEQ(*sp, radix, len)) {
530 *sp += len;
531 RESTORE_LC_NUMERIC();
532 return TRUE;
533 }
21431899
KW
534 }
535 RESTORE_LC_NUMERIC();
98994639
HS
536 }
537 /* always try "." if numeric radix didn't match because
538 * we may have data from different locales mixed */
539#endif
7918f24d
NC
540
541 PERL_ARGS_ASSERT_GROK_NUMERIC_RADIX;
542
98994639
HS
543 if (*sp < send && **sp == '.') {
544 ++*sp;
545 return TRUE;
546 }
547 return FALSE;
548}
549
550/*
3f7602fa 551=for apidoc grok_number_flags
98994639
HS
552
553Recognise (or not) a number. The type of the number is returned
554(0 if unrecognised), otherwise it is a bit-ORed combination of
555IS_NUMBER_IN_UV, IS_NUMBER_GREATER_THAN_UV_MAX, IS_NUMBER_NOT_INT,
aa8b85de 556IS_NUMBER_NEG, IS_NUMBER_INFINITY, IS_NUMBER_NAN (defined in perl.h).
60939fb8 557
cd164854 558If the value of the number can fit in a UV, it is returned in the *valuep
60939fb8
NC
559IS_NUMBER_IN_UV will be set to indicate that *valuep is valid, IS_NUMBER_IN_UV
560will never be set unless *valuep is valid, but *valuep may have been assigned
561to during processing even though IS_NUMBER_IN_UV is not set on return.
562If valuep is NULL, IS_NUMBER_IN_UV will be set for the same cases as when
563valuep is non-NULL, but no actual assignment (or SEGV) will occur.
564
565IS_NUMBER_NOT_INT will be set with IS_NUMBER_IN_UV if trailing decimals were
566seen (in which case *valuep gives the true value truncated to an integer), and
567IS_NUMBER_NEG if the number is negative (in which case *valuep holds the
568absolute value). IS_NUMBER_IN_UV is not set if e notation was used or the
569number is larger than a UV.
98994639 570
3f7602fa
TC
571C<flags> allows only C<PERL_SCAN_TRAILING>, which allows for trailing
572non-numeric text on an otherwise successful I<grok>, setting
573C<IS_NUMBER_TRAILING> on the result.
574
575=for apidoc grok_number
576
577Identical to grok_number_flags() with flags set to zero.
578
98994639
HS
579=cut
580 */
581int
582Perl_grok_number(pTHX_ const char *pv, STRLEN len, UV *valuep)
583{
3f7602fa
TC
584 PERL_ARGS_ASSERT_GROK_NUMBER;
585
586 return grok_number_flags(pv, len, valuep, 0);
587}
588
ff4eb398
JH
589/*
590=for apidoc grok_infnan
591
592Helper for grok_number(), accepts various ways of spelling "infinity"
593or "not a number", and returns one of the following flag combinations:
594
595 IS_NUMBER_INFINITE
596 IS_NUMBER_NAN
597 IS_NUMBER_INFINITE | IS_NUMBER_NEG
598 IS_NUMBER_NAN | IS_NUMBER_NEG
599 0
600
601If an infinity or not-a-number is recognized, the *sp will point to
602one past the end of the recognized string. If the recognition fails,
603zero is returned, and the *sp will not move.
604
605=cut
606*/
607
608int
609Perl_grok_infnan(const char** sp, const char* send)
610{
611 const char* s = *sp;
612 int flags = 0;
613
614 PERL_ARGS_ASSERT_GROK_INFNAN;
615
8c12dc63
JH
616 if (*s == '+') {
617 s++; if (s == send) return 0;
618 }
619 else if (*s == '-') {
ff4eb398
JH
620 flags |= IS_NUMBER_NEG; /* Yes, -NaN happens. Incorrect but happens. */
621 s++; if (s == send) return 0;
622 }
623
624 if (*s == '1') {
625 /* Visual C: 1.#SNAN, -1.#QNAN, 1#INF, 1#.IND (maybe also 1.#NAN) */
626 s++; if (s == send) return 0;
627 if (*s == '.') {
628 s++; if (s == send) return 0;
629 }
630 if (*s == '#') {
631 s++; if (s == send) return 0;
632 } else
633 return 0;
634 }
635
305b8651 636 if (isALPHA_FOLD_EQ(*s, 'I')) {
ff4eb398 637 /* INF or IND (1.#IND is indeterminate, a certain type of NAN) */
305b8651 638 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
ff4eb398 639 s++; if (s == send) return 0;
305b8651 640 if (isALPHA_FOLD_EQ(*s, 'F')) {
ff4eb398 641 s++;
305b8651
KW
642 if (s < send && (isALPHA_FOLD_EQ(*s, 'I'))) {
643 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
644 s++; if (s == send || isALPHA_FOLD_NE(*s, 'I')) return 0;
645 s++; if (s == send || isALPHA_FOLD_NE(*s, 'T')) return 0;
8c12dc63
JH
646 s++; if (s == send ||
647 /* allow either Infinity or Infinite */
0ec38c0a
JH
648 !(isALPHA_FOLD_EQ(*s, 'Y') ||
649 isALPHA_FOLD_EQ(*s, 'E'))) return 0;
650 s++; if (s < send) return 0;
ff4eb398
JH
651 } else if (*s)
652 return 0;
653 flags |= IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT;
654 }
305b8651 655 else if (isALPHA_FOLD_EQ(*s, 'D')) {
ff4eb398
JH
656 s++;
657 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
658 } else
659 return 0;
ff4eb398
JH
660 }
661 else {
662 /* NAN */
305b8651 663 if (isALPHA_FOLD_EQ(*s, 'S') || isALPHA_FOLD_EQ(*s, 'Q')) {
ff4eb398
JH
664 /* snan, qNaN */
665 /* XXX do something with the snan/qnan difference */
ae776a2c 666 s++; if (s == send) return 0;
ff4eb398
JH
667 }
668
305b8651
KW
669 if (isALPHA_FOLD_EQ(*s, 'N')) {
670 s++; if (s == send || isALPHA_FOLD_NE(*s, 'A')) return 0;
671 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
ff4eb398
JH
672 s++;
673
ae776a2c
JH
674 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
675
8c12dc63
JH
676 /* NaN can be followed by various stuff (NaNQ, NaNS), but
677 * there are also multiple different NaN values, and some
678 * implementations output the "payload" values,
679 * e.g. NaN123, NAN(abc), while some implementations just
680 * have weird stuff like NaN%. */
ae776a2c 681 s = send;
ff4eb398 682 }
ae776a2c
JH
683 else
684 return 0;
ff4eb398
JH
685 }
686
a1fe7cea
JH
687 *sp = s;
688 return flags;
ff4eb398
JH
689}
690
945b524a
JH
691static const UV uv_max_div_10 = UV_MAX / 10;
692static const U8 uv_max_mod_10 = UV_MAX % 10;
693
3f7602fa
TC
694int
695Perl_grok_number_flags(pTHX_ const char *pv, STRLEN len, UV *valuep, U32 flags)
696{
60939fb8 697 const char *s = pv;
c4420975 698 const char * const send = pv + len;
ae776a2c 699 const char *d;
60939fb8 700 int numtype = 0;
60939fb8 701
3f7602fa 702 PERL_ARGS_ASSERT_GROK_NUMBER_FLAGS;
7918f24d 703
60939fb8
NC
704 while (s < send && isSPACE(*s))
705 s++;
706 if (s == send) {
707 return 0;
708 } else if (*s == '-') {
709 s++;
710 numtype = IS_NUMBER_NEG;
711 }
712 else if (*s == '+')
aa42a541 713 s++;
60939fb8
NC
714
715 if (s == send)
716 return 0;
717
ae776a2c 718 /* The first digit (after optional sign): note that might
8c12dc63 719 * also point to "infinity" or "nan", or "1.#INF". */
ae776a2c
JH
720 d = s;
721
8c12dc63 722 /* next must be digit or the radix separator or beginning of infinity/nan */
60939fb8
NC
723 if (isDIGIT(*s)) {
724 /* UVs are at least 32 bits, so the first 9 decimal digits cannot
725 overflow. */
726 UV value = *s - '0';
727 /* This construction seems to be more optimiser friendly.
728 (without it gcc does the isDIGIT test and the *s - '0' separately)
729 With it gcc on arm is managing 6 instructions (6 cycles) per digit.
730 In theory the optimiser could deduce how far to unroll the loop
731 before checking for overflow. */
58bb9ec3
NC
732 if (++s < send) {
733 int digit = *s - '0';
60939fb8
NC
734 if (digit >= 0 && digit <= 9) {
735 value = value * 10 + digit;
58bb9ec3
NC
736 if (++s < send) {
737 digit = *s - '0';
60939fb8
NC
738 if (digit >= 0 && digit <= 9) {
739 value = value * 10 + digit;
58bb9ec3
NC
740 if (++s < send) {
741 digit = *s - '0';
60939fb8
NC
742 if (digit >= 0 && digit <= 9) {
743 value = value * 10 + digit;
58bb9ec3
NC
744 if (++s < send) {
745 digit = *s - '0';
60939fb8
NC
746 if (digit >= 0 && digit <= 9) {
747 value = value * 10 + digit;
58bb9ec3
NC
748 if (++s < send) {
749 digit = *s - '0';
60939fb8
NC
750 if (digit >= 0 && digit <= 9) {
751 value = value * 10 + digit;
58bb9ec3
NC
752 if (++s < send) {
753 digit = *s - '0';
60939fb8
NC
754 if (digit >= 0 && digit <= 9) {
755 value = value * 10 + digit;
58bb9ec3
NC
756 if (++s < send) {
757 digit = *s - '0';
60939fb8
NC
758 if (digit >= 0 && digit <= 9) {
759 value = value * 10 + digit;
58bb9ec3
NC
760 if (++s < send) {
761 digit = *s - '0';
60939fb8
NC
762 if (digit >= 0 && digit <= 9) {
763 value = value * 10 + digit;
58bb9ec3 764 if (++s < send) {
60939fb8
NC
765 /* Now got 9 digits, so need to check
766 each time for overflow. */
58bb9ec3 767 digit = *s - '0';
60939fb8 768 while (digit >= 0 && digit <= 9
945b524a
JH
769 && (value < uv_max_div_10
770 || (value == uv_max_div_10
771 && digit <= uv_max_mod_10))) {
60939fb8 772 value = value * 10 + digit;
58bb9ec3
NC
773 if (++s < send)
774 digit = *s - '0';
60939fb8
NC
775 else
776 break;
777 }
778 if (digit >= 0 && digit <= 9
51bd16da 779 && (s < send)) {
60939fb8
NC
780 /* value overflowed.
781 skip the remaining digits, don't
782 worry about setting *valuep. */
783 do {
784 s++;
785 } while (s < send && isDIGIT(*s));
786 numtype |=
787 IS_NUMBER_GREATER_THAN_UV_MAX;
788 goto skip_value;
789 }
790 }
791 }
98994639 792 }
60939fb8
NC
793 }
794 }
795 }
796 }
797 }
798 }
799 }
800 }
801 }
802 }
803 }
98994639 804 }
60939fb8 805 }
98994639 806 }
60939fb8
NC
807 numtype |= IS_NUMBER_IN_UV;
808 if (valuep)
809 *valuep = value;
810
811 skip_value:
812 if (GROK_NUMERIC_RADIX(&s, send)) {
813 numtype |= IS_NUMBER_NOT_INT;
814 while (s < send && isDIGIT(*s)) /* optional digits after the radix */
815 s++;
98994639 816 }
60939fb8
NC
817 }
818 else if (GROK_NUMERIC_RADIX(&s, send)) {
819 numtype |= IS_NUMBER_NOT_INT | IS_NUMBER_IN_UV; /* valuep assigned below */
820 /* no digits before the radix means we need digits after it */
821 if (s < send && isDIGIT(*s)) {
822 do {
823 s++;
824 } while (s < send && isDIGIT(*s));
825 if (valuep) {
826 /* integer approximation is valid - it's 0. */
827 *valuep = 0;
828 }
98994639 829 }
60939fb8 830 else
ae776a2c 831 return 0;
ff4eb398 832 }
60939fb8 833
926f5fc6 834 if (s > d && s < send) {
60939fb8 835 /* we can have an optional exponent part */
305b8651 836 if (isALPHA_FOLD_EQ(*s, 'e')) {
60939fb8
NC
837 s++;
838 if (s < send && (*s == '-' || *s == '+'))
839 s++;
840 if (s < send && isDIGIT(*s)) {
841 do {
842 s++;
843 } while (s < send && isDIGIT(*s));
844 }
3f7602fa
TC
845 else if (flags & PERL_SCAN_TRAILING)
846 return numtype | IS_NUMBER_TRAILING;
60939fb8 847 else
3f7602fa
TC
848 return 0;
849
850 /* The only flag we keep is sign. Blow away any "it's UV" */
851 numtype &= IS_NUMBER_NEG;
852 numtype |= IS_NUMBER_NOT_INT;
60939fb8
NC
853 }
854 }
855 while (s < send && isSPACE(*s))
856 s++;
857 if (s >= send)
aa8b85de 858 return numtype;
60939fb8
NC
859 if (len == 10 && memEQ(pv, "0 but true", 10)) {
860 if (valuep)
861 *valuep = 0;
862 return IS_NUMBER_IN_UV;
863 }
8c12dc63
JH
864 /* We could be e.g. at "Inf" or "NaN", or at the "#" of "1.#INF". */
865 if ((s + 2 < send) && strchr("inqs#", toFOLD(*s))) {
866 /* Really detect inf/nan. Start at d, not s, since the above
867 * code might have already consumed the "1." or "1". */
868 int infnan = Perl_grok_infnan(&d, send);
869 if ((infnan & IS_NUMBER_INFINITY)) {
870 return (numtype | infnan); /* Keep sign for infinity. */
871 }
872 else if ((infnan & IS_NUMBER_NAN)) {
873 return (numtype | infnan) & ~IS_NUMBER_NEG; /* Clear sign for nan. */
874 }
875 }
3f7602fa
TC
876 else if (flags & PERL_SCAN_TRAILING) {
877 return numtype | IS_NUMBER_TRAILING;
878 }
879
60939fb8 880 return 0;
98994639
HS
881}
882
6313e544 883/*
d62b8c6a 884=for apidoc grok_atou
6313e544 885
d62b8c6a 886grok_atou is a safer replacement for atoi and strtol.
6313e544 887
d62b8c6a
JH
888grok_atou parses a C-style zero-byte terminated string, looking for
889a decimal unsigned integer.
338aa8b0 890
d62b8c6a
JH
891Returns the unsigned integer, if a valid value can be parsed
892from the beginning of the string.
f4379102 893
d62b8c6a 894Accepts only the decimal digits '0'..'9'.
6313e544 895
d62b8c6a
JH
896As opposed to atoi or strtol, grok_atou does NOT allow optional
897leading whitespace, or negative inputs. If such features are
898required, the calling code needs to explicitly implement those.
6313e544 899
d62b8c6a 900If a valid value cannot be parsed, returns either zero (if non-digits
75feedba 901are met before any digits) or UV_MAX (if the value overflows).
6313e544 902
d62b8c6a
JH
903Note that extraneous leading zeros also count as an overflow
904(meaning that only "0" is the zero).
338aa8b0 905
d62b8c6a 906On failure, the *endptr is also set to NULL, unless endptr is NULL.
338aa8b0
JH
907
908Trailing non-digit bytes are allowed if the endptr is non-NULL.
6313e544
JH
909On return the *endptr will contain the pointer to the first non-digit byte.
910
6313e544 911If the endptr is NULL, the first non-digit byte MUST be
f4379102 912the zero byte terminating the pv, or zero will be returned.
6313e544 913
d62b8c6a
JH
914Background: atoi has severe problems with illegal inputs, it cannot be
915used for incremental parsing, and therefore should be avoided
916atoi and strtol are also affected by locale settings, which can also be
917seen as a bug (global state controlled by user environment).
918
6313e544
JH
919=cut
920*/
921
75feedba 922UV
6313e544
JH
923Perl_grok_atou(const char *pv, const char** endptr)
924{
925 const char* s = pv;
926 const char** eptr;
927 const char* end2; /* Used in case endptr is NULL. */
75feedba 928 UV val = 0; /* The return value. */
6313e544
JH
929
930 PERL_ARGS_ASSERT_GROK_ATOU;
931
932 eptr = endptr ? endptr : &end2;
75feedba
JH
933 if (isDIGIT(*s)) {
934 /* Single-digit inputs are quite common. */
6313e544 935 val = *s++ - '0';
75feedba
JH
936 if (isDIGIT(*s)) {
937 /* Extra leading zeros cause overflow. */
938 if (val == 0) {
939 *eptr = NULL;
940 return UV_MAX;
941 }
942 while (isDIGIT(*s)) {
943 /* This could be unrolled like in grok_number(), but
944 * the expected uses of this are not speed-needy, and
945 * unlikely to need full 64-bitness. */
946 U8 digit = *s++ - '0';
945b524a
JH
947 if (val < uv_max_div_10 ||
948 (val == uv_max_div_10 && digit <= uv_max_mod_10)) {
75feedba
JH
949 val = val * 10 + digit;
950 } else {
6313e544 951 *eptr = NULL;
75feedba 952 return UV_MAX;
6313e544 953 }
6313e544
JH
954 }
955 }
75feedba
JH
956 }
957 if (s == pv) {
958 *eptr = NULL; /* If no progress, failed to parse anything. */
959 return 0;
6313e544
JH
960 }
961 if (endptr == NULL && *s) {
962 return 0; /* If endptr is NULL, no trailing non-digits allowed. */
963 }
964 *eptr = s;
965 return val;
966}
967
a4eca1d4 968#ifndef USE_QUADMATH
4801ca72 969STATIC NV
98994639
HS
970S_mulexp10(NV value, I32 exponent)
971{
972 NV result = 1.0;
973 NV power = 10.0;
974 bool negative = 0;
975 I32 bit;
976
977 if (exponent == 0)
978 return value;
659c4b96
DM
979 if (value == 0)
980 return (NV)0;
87032ba1 981
24866caa 982 /* On OpenVMS VAX we by default use the D_FLOAT double format,
67597c89 983 * and that format does not have *easy* capabilities [1] for
24866caa
CB
984 * overflowing doubles 'silently' as IEEE fp does. We also need
985 * to support G_FLOAT on both VAX and Alpha, and though the exponent
986 * range is much larger than D_FLOAT it still doesn't do silent
987 * overflow. Therefore we need to detect early whether we would
988 * overflow (this is the behaviour of the native string-to-float
989 * conversion routines, and therefore of native applications, too).
67597c89 990 *
24866caa
CB
991 * [1] Trying to establish a condition handler to trap floating point
992 * exceptions is not a good idea. */
87032ba1
JH
993
994 /* In UNICOS and in certain Cray models (such as T90) there is no
995 * IEEE fp, and no way at all from C to catch fp overflows gracefully.
996 * There is something you can do if you are willing to use some
997 * inline assembler: the instruction is called DFI-- but that will
998 * disable *all* floating point interrupts, a little bit too large
999 * a hammer. Therefore we need to catch potential overflows before
1000 * it's too late. */
353813d9 1001
85bba25f 1002#if ((defined(VMS) && !defined(_IEEE_FP)) || defined(_UNICOS)) && defined(NV_MAX_10_EXP)
353813d9 1003 STMT_START {
c4420975 1004 const NV exp_v = log10(value);
353813d9
HS
1005 if (exponent >= NV_MAX_10_EXP || exponent + exp_v >= NV_MAX_10_EXP)
1006 return NV_MAX;
1007 if (exponent < 0) {
1008 if (-(exponent + exp_v) >= NV_MAX_10_EXP)
1009 return 0.0;
1010 while (-exponent >= NV_MAX_10_EXP) {
1011 /* combination does not overflow, but 10^(-exponent) does */
1012 value /= 10;
1013 ++exponent;
1014 }
1015 }
1016 } STMT_END;
87032ba1
JH
1017#endif
1018
353813d9
HS
1019 if (exponent < 0) {
1020 negative = 1;
1021 exponent = -exponent;
b27804d8
DM
1022#ifdef NV_MAX_10_EXP
1023 /* for something like 1234 x 10^-309, the action of calculating
1024 * the intermediate value 10^309 then returning 1234 / (10^309)
1025 * will fail, since 10^309 becomes infinity. In this case try to
1026 * refactor it as 123 / (10^308) etc.
1027 */
1028 while (value && exponent > NV_MAX_10_EXP) {
1029 exponent--;
1030 value /= 10;
1031 }
1032#endif
353813d9 1033 }
98994639
HS
1034 for (bit = 1; exponent; bit <<= 1) {
1035 if (exponent & bit) {
1036 exponent ^= bit;
1037 result *= power;
236f0012
CB
1038 /* Floating point exceptions are supposed to be turned off,
1039 * but if we're obviously done, don't risk another iteration.
1040 */
1041 if (exponent == 0) break;
98994639
HS
1042 }
1043 power *= power;
1044 }
1045 return negative ? value / result : value * result;
1046}
a4eca1d4 1047#endif /* #ifndef USE_QUADMATH */
98994639
HS
1048
1049NV
1050Perl_my_atof(pTHX_ const char* s)
1051{
1052 NV x = 0.0;
a4eca1d4
JH
1053#ifdef USE_QUADMATH
1054 Perl_my_atof2(aTHX_ s, &x);
1055 return x;
1056#else
1057# ifdef USE_LOCALE_NUMERIC
7918f24d
NC
1058 PERL_ARGS_ASSERT_MY_ATOF;
1059
a2287a13
KW
1060 {
1061 DECLARE_STORE_LC_NUMERIC_SET_TO_NEEDED();
d6ded950 1062 if (PL_numeric_radix_sv && IN_LC(LC_NUMERIC)) {
e4850248
KW
1063 const char *standard = NULL, *local = NULL;
1064 bool use_standard_radix;
98994639 1065
e4850248
KW
1066 /* Look through the string for the first thing that looks like a
1067 * decimal point: either the value in the current locale or the
1068 * standard fallback of '.'. The one which appears earliest in the
1069 * input string is the one that we should have atof look for. Note
1070 * that we have to determine this beforehand because on some
1071 * systems, Perl_atof2 is just a wrapper around the system's atof.
1072 * */
1073 standard = strchr(s, '.');
1074 local = strstr(s, SvPV_nolen(PL_numeric_radix_sv));
78787052 1075
e4850248 1076 use_standard_radix = standard && (!local || standard < local);
78787052 1077
e4850248
KW
1078 if (use_standard_radix)
1079 SET_NUMERIC_STANDARD();
78787052 1080
e4850248 1081 Perl_atof2(s, x);
78787052 1082
e4850248
KW
1083 if (use_standard_radix)
1084 SET_NUMERIC_LOCAL();
1085 }
1086 else
1087 Perl_atof2(s, x);
a2287a13
KW
1088 RESTORE_LC_NUMERIC();
1089 }
a4eca1d4 1090# else
a36244b7 1091 Perl_atof2(s, x);
a4eca1d4 1092# endif
98994639
HS
1093#endif
1094 return x;
1095}
1096
829757a4
JH
1097static char*
1098S_my_atof_infnan(const char* s, bool negative, const char* send, NV* value)
1099{
1100 const char *p0 = negative ? s - 1 : s;
1101 const char *p = p0;
1102 int infnan = grok_infnan(&p, send);
1103 if (infnan && p != p0) {
1104 /* If we can generate inf/nan directly, let's do so. */
1105#ifdef NV_INF
1106 if ((infnan & IS_NUMBER_INFINITY)) {
1107 *value = (infnan & IS_NUMBER_NEG) ? -NV_INF: NV_INF;
1108 return (char*)p;
1109 }
1110#endif
1111#ifdef NV_NAN
1112 if ((infnan & IS_NUMBER_NAN)) {
1113 *value = NV_NAN;
1114 return (char*)p;
1115 }
1116#endif
1117#ifdef Perl_strtod
68611e6f 1118 /* If still here, we didn't have either NV_INF or NV_NAN,
829757a4
JH
1119 * and can try falling back to native strtod/strtold.
1120 *
68611e6f
JH
1121 * (Though, are our NV_INF or NV_NAN ever not defined?)
1122 *
829757a4
JH
1123 * The native interface might not recognize all the possible
1124 * inf/nan strings Perl recognizes. What we can try
1125 * is to try faking the input. We will try inf/-inf/nan
1126 * as the most promising/portable input. */
1127 {
1128 const char* fake = NULL;
1129 char* endp;
1130 NV nv;
1131 if ((infnan & IS_NUMBER_INFINITY)) {
1132 fake = ((infnan & IS_NUMBER_NEG)) ? "-inf" : "inf";
1133 }
1134 else if ((infnan & IS_NUMBER_NAN)) {
1135 fake = "nan";
1136 }
1137 assert(fake);
1138 nv = Perl_strtod(fake, &endp);
1139 if (fake != endp) {
1140 if ((infnan & IS_NUMBER_INFINITY)) {
1141#ifdef Perl_isinf
1142 if (Perl_isinf(nv))
1143 *value = nv;
1144#else
1145 /* last resort, may generate SIGFPE */
1146 *value = Perl_exp((NV)1e9);
1147 if ((infnan & IS_NUMBER_NEG))
1148 *value = -*value;
1149#endif
1150 return (char*)p; /* p, not endp */
1151 }
1152 else if ((infnan & IS_NUMBER_NAN)) {
1153#ifdef Perl_isnan
1154 if (Perl_isnan(nv))
1155 *value = nv;
1156#else
1157 /* last resort, may generate SIGFPE */
1158 *value = Perl_log((NV)-1.0);
1159#endif
1160 return (char*)p; /* p, not endp */
1161 }
1162 }
1163 }
1164#endif /* #ifdef Perl_strtod */
1165 }
1166 return NULL;
1167}
1168
98994639
HS
1169char*
1170Perl_my_atof2(pTHX_ const char* orig, NV* value)
1171{
e1ec3a88 1172 const char* s = orig;
a4eca1d4
JH
1173 NV result[3] = {0.0, 0.0, 0.0};
1174#if defined(USE_PERL_ATOF) || defined(USE_QUADMATH)
ae776a2c 1175 const char* send = s + strlen(orig); /* one past the last */
a4eca1d4
JH
1176 bool negative = 0;
1177#endif
1178#if defined(USE_PERL_ATOF) && !defined(USE_QUADMATH)
1179 UV accumulator[2] = {0,0}; /* before/after dp */
8194bf88 1180 bool seen_digit = 0;
20f6aaab
AS
1181 I32 exp_adjust[2] = {0,0};
1182 I32 exp_acc[2] = {-1, -1};
1183 /* the current exponent adjust for the accumulators */
98994639 1184 I32 exponent = 0;
8194bf88 1185 I32 seen_dp = 0;
20f6aaab
AS
1186 I32 digit = 0;
1187 I32 old_digit = 0;
8194bf88 1188 I32 sig_digits = 0; /* noof significant digits seen so far */
a4eca1d4 1189#endif
8194bf88 1190
a4eca1d4 1191#if defined(USE_PERL_ATOF) || defined(USE_QUADMATH)
7918f24d
NC
1192 PERL_ARGS_ASSERT_MY_ATOF2;
1193
a4eca1d4
JH
1194 /* leading whitespace */
1195 while (isSPACE(*s))
1196 ++s;
1197
1198 /* sign */
1199 switch (*s) {
1200 case '-':
1201 negative = 1;
1202 /* FALLTHROUGH */
1203 case '+':
1204 ++s;
1205 }
1206#endif
1207
1208#ifdef USE_QUADMATH
1209 {
1210 char* endp;
1211 if ((endp = S_my_atof_infnan(s, negative, send, value)))
1212 return endp;
1213 result[2] = strtoflt128(s, &endp);
1214 if (s != endp) {
1215 *value = negative ? -result[2] : result[2];
1216 return endp;
1217 }
1218 return NULL;
1219 }
1220#elif defined(USE_PERL_ATOF)
1221
8194bf88
DM
1222/* There is no point in processing more significant digits
1223 * than the NV can hold. Note that NV_DIG is a lower-bound value,
1224 * while we need an upper-bound value. We add 2 to account for this;
1225 * since it will have been conservative on both the first and last digit.
1226 * For example a 32-bit mantissa with an exponent of 4 would have
1227 * exact values in the set
1228 * 4
1229 * 8
1230 * ..
1231 * 17179869172
1232 * 17179869176
1233 * 17179869180
1234 *
1235 * where for the purposes of calculating NV_DIG we would have to discount
1236 * both the first and last digit, since neither can hold all values from
1237 * 0..9; but for calculating the value we must examine those two digits.
1238 */
ffa277e5
EAS
1239#ifdef MAX_SIG_DIG_PLUS
1240 /* It is not necessarily the case that adding 2 to NV_DIG gets all the
1241 possible digits in a NV, especially if NVs are not IEEE compliant
1242 (e.g., long doubles on IRIX) - Allen <allens@cpan.org> */
1243# define MAX_SIG_DIGITS (NV_DIG+MAX_SIG_DIG_PLUS)
1244#else
1245# define MAX_SIG_DIGITS (NV_DIG+2)
1246#endif
8194bf88
DM
1247
1248/* the max number we can accumulate in a UV, and still safely do 10*N+9 */
1249#define MAX_ACCUMULATE ( (UV) ((UV_MAX - 9)/10))
98994639 1250
ae776a2c 1251 {
829757a4
JH
1252 const char* endp;
1253 if ((endp = S_my_atof_infnan(s, negative, send, value)))
1254 return (char*)endp;
ae776a2c 1255 }
2b54f59f 1256
8194bf88
DM
1257 /* we accumulate digits into an integer; when this becomes too
1258 * large, we add the total to NV and start again */
98994639 1259
8194bf88
DM
1260 while (1) {
1261 if (isDIGIT(*s)) {
1262 seen_digit = 1;
20f6aaab 1263 old_digit = digit;
8194bf88 1264 digit = *s++ - '0';
20f6aaab
AS
1265 if (seen_dp)
1266 exp_adjust[1]++;
98994639 1267
8194bf88
DM
1268 /* don't start counting until we see the first significant
1269 * digit, eg the 5 in 0.00005... */
1270 if (!sig_digits && digit == 0)
1271 continue;
1272
1273 if (++sig_digits > MAX_SIG_DIGITS) {
98994639 1274 /* limits of precision reached */
20f6aaab
AS
1275 if (digit > 5) {
1276 ++accumulator[seen_dp];
1277 } else if (digit == 5) {
1278 if (old_digit % 2) { /* round to even - Allen */
1279 ++accumulator[seen_dp];
1280 }
1281 }
1282 if (seen_dp) {
1283 exp_adjust[1]--;
1284 } else {
1285 exp_adjust[0]++;
1286 }
8194bf88 1287 /* skip remaining digits */
98994639 1288 while (isDIGIT(*s)) {
98994639 1289 ++s;
20f6aaab
AS
1290 if (! seen_dp) {
1291 exp_adjust[0]++;
1292 }
98994639
HS
1293 }
1294 /* warn of loss of precision? */
98994639 1295 }
8194bf88 1296 else {
20f6aaab 1297 if (accumulator[seen_dp] > MAX_ACCUMULATE) {
8194bf88 1298 /* add accumulator to result and start again */
20f6aaab
AS
1299 result[seen_dp] = S_mulexp10(result[seen_dp],
1300 exp_acc[seen_dp])
1301 + (NV)accumulator[seen_dp];
1302 accumulator[seen_dp] = 0;
1303 exp_acc[seen_dp] = 0;
98994639 1304 }
20f6aaab
AS
1305 accumulator[seen_dp] = accumulator[seen_dp] * 10 + digit;
1306 ++exp_acc[seen_dp];
98994639 1307 }
8194bf88 1308 }
e1ec3a88 1309 else if (!seen_dp && GROK_NUMERIC_RADIX(&s, send)) {
8194bf88 1310 seen_dp = 1;
20f6aaab 1311 if (sig_digits > MAX_SIG_DIGITS) {
c86f7df5 1312 do {
20f6aaab 1313 ++s;
c86f7df5 1314 } while (isDIGIT(*s));
20f6aaab
AS
1315 break;
1316 }
8194bf88
DM
1317 }
1318 else {
1319 break;
98994639
HS
1320 }
1321 }
1322
20f6aaab
AS
1323 result[0] = S_mulexp10(result[0], exp_acc[0]) + (NV)accumulator[0];
1324 if (seen_dp) {
1325 result[1] = S_mulexp10(result[1], exp_acc[1]) + (NV)accumulator[1];
1326 }
98994639 1327
305b8651 1328 if (seen_digit && (isALPHA_FOLD_EQ(*s, 'e'))) {
98994639
HS
1329 bool expnegative = 0;
1330
1331 ++s;
1332 switch (*s) {
1333 case '-':
1334 expnegative = 1;
924ba076 1335 /* FALLTHROUGH */
98994639
HS
1336 case '+':
1337 ++s;
1338 }
1339 while (isDIGIT(*s))
1340 exponent = exponent * 10 + (*s++ - '0');
1341 if (expnegative)
1342 exponent = -exponent;
1343 }
1344
20f6aaab
AS
1345
1346
98994639 1347 /* now apply the exponent */
20f6aaab
AS
1348
1349 if (seen_dp) {
1350 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0])
1351 + S_mulexp10(result[1],exponent-exp_adjust[1]);
1352 } else {
1353 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0]);
1354 }
98994639
HS
1355
1356 /* now apply the sign */
1357 if (negative)
20f6aaab 1358 result[2] = -result[2];
a36244b7 1359#endif /* USE_PERL_ATOF */
20f6aaab 1360 *value = result[2];
73d840c0 1361 return (char *)s;
98994639
HS
1362}
1363
5d34af89 1364/*
3d9d9213 1365=for apidoc isinfnan
5d34af89
JH
1366
1367Perl_isinfnan() is utility function that returns true if the NV
1368argument is either an infinity or a NaN, false otherwise. To test
1369in more detail, use Perl_isinf() and Perl_isnan().
1370
68611e6f
JH
1371This is also the logical inverse of Perl_isfinite().
1372
5d34af89
JH
1373=cut
1374*/
1cd88304
JH
1375bool
1376Perl_isinfnan(NV nv)
1377{
1378#ifdef Perl_isinf
1379 if (Perl_isinf(nv))
1380 return TRUE;
1381#endif
1382#ifdef Perl_isnan
1383 if (Perl_isnan(nv))
1384 return TRUE;
1385#endif
1386 return FALSE;
1387}
1388
d67dac15 1389#ifndef HAS_MODFL
68611e6f
JH
1390/* C99 has truncl, pre-C99 Solaris had aintl. We can use either with
1391 * copysignl to emulate modfl, which is in some platforms missing or
1392 * broken. */
d67dac15
JH
1393# if defined(HAS_TRUNCL) && defined(HAS_COPYSIGNL)
1394long double
1395Perl_my_modfl(long double x, long double *ip)
1396{
68611e6f
JH
1397 *ip = truncl(x);
1398 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
d67dac15
JH
1399}
1400# elif defined(HAS_AINTL) && defined(HAS_COPYSIGNL)
55954f19
JH
1401long double
1402Perl_my_modfl(long double x, long double *ip)
1403{
68611e6f
JH
1404 *ip = aintl(x);
1405 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
55954f19 1406}
d67dac15 1407# endif
55954f19
JH
1408#endif
1409
68611e6f 1410/* Similarly, with ilobl and scalbnl we can emulate frexpl. */
55954f19
JH
1411#if ! defined(HAS_FREXPL) && defined(HAS_ILOGBL) && defined(HAS_SCALBNL)
1412long double
1413Perl_my_frexpl(long double x, int *e) {
68611e6f
JH
1414 *e = x == 0.0L ? 0 : ilogbl(x) + 1;
1415 return (scalbnl(x, -*e));
55954f19
JH
1416}
1417#endif
66610fdd
RGS
1418
1419/*
ed140128
AD
1420=for apidoc Perl_signbit
1421
1422Return a non-zero integer if the sign bit on an NV is set, and 0 if
1423it is not.
1424
1425If Configure detects this system has a signbit() that will work with
1426our NVs, then we just use it via the #define in perl.h. Otherwise,
8b7fad81
JH
1427fall back on this implementation. The main use of this function
1428is catching -0.0.
ed140128
AD
1429
1430Configure notes: This function is called 'Perl_signbit' instead of a
1431plain 'signbit' because it is easy to imagine a system having a signbit()
1432function or macro that doesn't happen to work with our particular choice
1433of NVs. We shouldn't just re-#define signbit as Perl_signbit and expect
1434the standard system headers to be happy. Also, this is a no-context
1435function (no pTHX_) because Perl_signbit() is usually re-#defined in
1436perl.h as a simple macro call to the system's signbit().
1437Users should just always call Perl_signbit().
1438
1439=cut
1440*/
1441#if !defined(HAS_SIGNBIT)
1442int
1443Perl_signbit(NV x) {
8b7fad81
JH
1444# ifdef Perl_fp_class_nzero
1445 if (x == 0)
1446 return Perl_fp_class_nzero(x);
8b7fad81 1447# endif
3585840c 1448 return (x < 0.0) ? 1 : 0;
ed140128
AD
1449}
1450#endif
1451
1452/*
66610fdd
RGS
1453 * Local variables:
1454 * c-indentation-style: bsd
1455 * c-basic-offset: 4
14d04a33 1456 * indent-tabs-mode: nil
66610fdd
RGS
1457 * End:
1458 *
14d04a33 1459 * ex: set ts=8 sts=4 sw=4 et:
37442d52 1460 */