This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
infnan: store the nan payload error in an optional SV
[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
30419b52
JH
550#if 0
551/* For debugging. */
552static void
553S_hexdump_nv(NV nv)
554{
555 int i;
556 /* Remember that NVSIZE may include garbage bytes, the most
557 * notable case being the x86 80-bit extended precision long doubles,
558 * which have 6 or 2 unused bytes (NVSIZE = 16 or NVSIZE = 12). */
559 for (i = 0; i < NVSIZE; i++) {
560 PerlIO_printf(Perl_debug_log, "%02x ", ((U8*)&nv)[i]);
561 }
562 PerlIO_printf(Perl_debug_log, "\n");
563}
29b62199
JH
564#endif
565
98994639 566/*
b74dc0b3
JH
567=for apidoc nan_hibyte
568
569Given an NV, returns pointer to the byte containing the most
570significant bit of the NaN, this bit is most commonly the
571quiet/signaling bit of the NaN. The mask will contain a mask
572appropriate for manipulating the most significant bit.
573Note that this bit may not be the highest bit of the byte.
574
575If the NV is not a NaN, returns NULL.
576
577Most platforms have "high bit is one" -> quiet nan.
578The known opposite exceptions are older MIPS and HPPA platforms.
579
580Some platforms do not differentiate between quiet and signaling NaNs.
581
582=cut
583*/
584U8*
585Perl_nan_hibyte(NV *nvp, U8* mask)
586{
587 STRLEN i = (NV_MANT_REAL_DIG - 1) / 8;
b74dc0b3
JH
588
589 PERL_ARGS_ASSERT_NAN_HIBYTE;
590
29b62199
JH
591#if defined(USE_LONG_DOUBLE) && (LONG_DOUBLEKIND == LONG_DOUBLE_IS_X86_80_BIT_LITTLE_ENDIAN)
592 /* See the definition of NV_NAN_BITS. */
593 *mask = 1 << 6;
594#else
595 {
596 STRLEN j = (NV_MANT_REAL_DIG - 1) % 8;
597 *mask = 1 << j;
598 }
599#endif
b74dc0b3
JH
600#ifdef NV_BIG_ENDIAN
601 return (U8*) nvp + NVSIZE - 1 - i;
602#endif
603#ifdef NV_LITTLE_ENDIAN
604 return (U8*) nvp + i;
605#endif
606}
607
608/*
6640aa2c
JH
609=for apidoc nan_signaling_set
610
611Set or unset the NaN signaling-ness.
612
613Of those platforms that differentiate between quiet and signaling
614platforms the majority has the semantics of the most significant bit
615being on meaning quiet NaN, so for signaling we need to clear the bit.
616
617Some platforms (older MIPS, and HPPA) have the opposite
618semantics, and we set the bit for a signaling NaN.
619
620=cut
621*/
622void
be181dc9 623Perl_nan_signaling_set(pTHX_ NV *nvp, bool signaling)
6640aa2c
JH
624{
625 U8 mask;
626 U8* hibyte;
627
628 PERL_ARGS_ASSERT_NAN_SIGNALING_SET;
629
630 hibyte = nan_hibyte(nvp, &mask);
631 if (hibyte) {
632 const NV nan = NV_NAN;
30419b52
JH
633 /* Decent optimizers should make the irrelevant branch to disappear.
634 * XXX Configure scan */
6640aa2c
JH
635 if ((((U8*)&nan)[hibyte - (U8*)nvp] & mask)) {
636 /* x86 style: the most significant bit of the NaN is off
637 * for a signaling NaN, and on for a quiet NaN. */
638 if (signaling) {
639 *hibyte &= ~mask;
640 } else {
641 *hibyte |= mask;
642 }
643 } else {
644 /* MIPS/HPPA style: the most significant bit of the NaN is on
645 * for a signaling NaN, and off for a quiet NaN. */
646 if (signaling) {
647 *hibyte |= mask;
648 } else {
649 *hibyte &= ~mask;
650 }
651 }
652 }
653}
654
655/*
a307a0b0
JH
656=for apidoc nan_is_signaling
657
658Returns true if the nv is a NaN is a signaling NaN.
659
660=cut
661*/
662int
663Perl_nan_is_signaling(NV nv)
664{
665 /* Quiet NaN bit pattern (64-bit doubles, ignore endianness):
666 * x86 00 00 00 00 00 00 f8 7f
667 * sparc 7f ff ff ff ff ff ff ff
668 * mips 7f f7 ff ff ff ff ff ff
669 * hppa 7f f4 00 00 00 00 00 00
670 * The "7ff" is the exponent. The most significant bit of the NaN
671 * (note: here, not the most significant bit of the byte) is of
672 * interest: in the x86 style (also in sparc) the bit on means
30419b52 673 * 'quiet', in the mips/hppa style the bit off means 'quiet'. */
a307a0b0
JH
674#ifdef Perl_fp_classify_snan
675 return Perl_fp_classify_snan(nv);
676#else
677 if (Perl_isnan(nv)) {
678 U8 mask;
be181dc9 679 U8 *hibyte = nan_hibyte(&nv, &mask);
30419b52
JH
680 if (hibyte) {
681 /* Hoping NV_NAN is a quiet nan - this might be a false hope.
682 * XXX Configure test */
683 const NV nan = NV_NAN;
684 return (*hibyte & mask) != (((U8*)&nan)[hibyte - (U8*)&nv] & mask);
685 }
a307a0b0 686 }
30419b52 687 return 0;
a307a0b0
JH
688#endif
689}
690
4f89311d
JH
691/* The largest known floating point numbers are the IEEE quadruple
692 * precision of 128 bits. */
693#define MAX_NV_BYTES (128/8)
694
4f89311d
JH
695/*
696
697=for apidoc nan_payload_set
698
699Set the NaN payload of the nv.
700
701The first byte is the highest order byte of the payload (big-endian).
702
703The signaling flag, if true, turns the generated NaN into a signaling one.
704In most platforms this means turning _off_ the most significant bit of the
705NaN. Note the _most_ - some platforms have the opposite semantics.
706Do not assume any portability of the NaN semantics.
707
708=cut
709*/
710void
4258cf90 711Perl_nan_payload_set(pTHX_ NV *nvp, SV* svp, const void *bytes, STRLEN byten, bool signaling)
4f89311d
JH
712{
713 /* How many bits we can set in the payload.
714 *
715 * Note that whether the most signicant bit is a quiet or
716 * signaling NaN is actually unstandardized. Most platforms use
717 * it as the 'quiet' bit. The known exceptions to this are older
718 * MIPS, and HPPA.
719 *
720 * Yet another unstandardized area is what does the difference
721 * actually mean - if it exists: some platforms do not even have
722 * signaling NaNs.
723 *
724 * C99 nan() is supposed to generate quiet NaNs. */
29b62199 725 int bits = NV_NAN_BITS;
30419b52
JH
726 U8 mask;
727 U8* hibyte;
728 U8 hibit;
4f89311d
JH
729
730 STRLEN i, nvi;
eb254f26 731 bool overflow = FALSE;
4f89311d
JH
732
733 /* XXX None of this works for doubledouble platforms, or for mixendians. */
734
735 PERL_ARGS_ASSERT_NAN_PAYLOAD_SET;
736
737 *nvp = NV_NAN;
30419b52
JH
738 hibyte = nan_hibyte(nvp, &mask);
739 hibit = *hibyte & mask;
4f89311d
JH
740
741#ifdef NV_BIG_ENDIAN
742 nvi = NVSIZE - 1;
743#endif
744#ifdef NV_LITTLE_ENDIAN
745 nvi = 0;
746#endif
747
748 if (byten > MAX_NV_BYTES) {
749 byten = MAX_NV_BYTES;
eb254f26 750 overflow = TRUE;
4f89311d
JH
751 }
752 for (i = 0; bits > 0; i++) {
753 U8 b = i < byten ? ((U8*) bytes)[i] : 0;
754 if (bits > 0 && bits < 8) {
755 U8 m = (1 << bits) - 1;
756 ((U8*)nvp)[nvi] &= ~m;
757 ((U8*)nvp)[nvi] |= b & m;
758 bits = 0;
759 } else {
760 ((U8*)nvp)[nvi] = b;
761 bits -= 8;
762 }
763#ifdef NV_BIG_ENDIAN
764 nvi--;
765#endif
766#ifdef NV_LITTLE_ENDIAN
767 nvi++;
768#endif
769 }
30419b52
JH
770 if (hibit) {
771 *hibyte |= mask;
772 } else {
773 *hibyte &= ~mask;
774 }
eb254f26 775 if (overflow) {
4258cf90
JH
776 if (svp) {
777 sv_setpvf(svp, "NaN payload overflowed %d bits", NV_NAN_BITS);
778 }
4f89311d
JH
779 }
780 nan_signaling_set(nvp, signaling);
781}
782
a307a0b0 783/*
569f27e5
JH
784=for apidoc grok_nan_payload
785
786Helper for grok_nan().
787
788Parses the "..." in C99-style "nan(...)" strings, and sets the nvp accordingly.
789
790If you want the parse the "nan" part you need to use grok_nan().
791
792=cut
793*/
794const char *
4258cf90 795Perl_grok_nan_payload(pTHX_ const char* s, const char* send, bool signaling, int *flags, NV* nvp, SV* svp)
569f27e5
JH
796{
797 U8 bytes[MAX_NV_BYTES];
798 STRLEN byten = 0;
799 const char *t = send - 1; /* minus one for ')' */
eb254f26
JH
800 bool overflow = FALSE;
801 bool bogus = FALSE;
802 const char *orig = s;
569f27e5
JH
803
804 PERL_ARGS_ASSERT_GROK_NAN_PAYLOAD;
805
806 /* XXX: legacy nan payload formats like "nan123",
807 * "nan0xabc", or "nan(s123)" ("s" for signaling). */
808
809 while (t > s && isSPACE(*t)) t--;
99fcdd4d 810
569f27e5 811 if (*t != ')') {
99fcdd4d 812 U8 bytes[1] = { 0 };
4258cf90 813 nan_payload_set(nvp, svp, bytes, 1, signaling);
99fcdd4d 814 return t;
569f27e5
JH
815 }
816
817 if (++s == send) {
818 *flags |= IS_NUMBER_TRAILING;
819 return s;
820 }
821
822 while (s < t && byten < MAX_NV_BYTES) {
823 UV uv;
824 int nantype = 0;
825
826 if (s[0] == '0' && s + 2 < t &&
827 isALPHA_FOLD_EQ(s[1], 'x') &&
828 isXDIGIT(s[2])) {
829 const char *u = s + 3;
830 STRLEN len;
831 I32 uvflags;
832
833 while (isXDIGIT(*u)) u++;
834 len = u - s;
835 uvflags = PERL_SCAN_ALLOW_UNDERSCORES;
836 uv = grok_hex(s, &len, &uvflags, NULL);
837 if ((uvflags & PERL_SCAN_GREATER_THAN_UV_MAX)) {
838 nantype = 0;
839 } else {
840 nantype = IS_NUMBER_IN_UV;
841 }
842 s += len;
843 } else if (s[0] == '0' && s + 2 < t &&
844 isALPHA_FOLD_EQ(s[1], 'b') &&
845 (s[2] == '0' || s[2] == '1')) {
846 const char *u = s + 3;
847 STRLEN len;
848 I32 uvflags;
849
850 while (*u == '0' || *u == '1') u++;
851 len = u - s;
852 uvflags = PERL_SCAN_ALLOW_UNDERSCORES;
853 uv = grok_bin(s, &len, &uvflags, NULL);
854 if ((uvflags & PERL_SCAN_GREATER_THAN_UV_MAX)) {
855 nantype = 0;
856 } else {
857 nantype = IS_NUMBER_IN_UV;
858 }
859 s += len;
860 } else if ((s[0] == '\'' || s[0] == '"') &&
861 s + 2 < t && t[-1] == s[0]) {
862 /* Perl extension: if the input looks like a string
863 * constant ('' or ""), read its bytes as-they-come. */
864 STRLEN n = t - s - 2;
865 STRLEN i;
866 if ((n > MAX_NV_BYTES - byten) ||
867 (n * 8 > NV_MANT_REAL_DIG)) {
eb254f26 868 overflow = TRUE;
569f27e5
JH
869 break;
870 }
871 /* Copy the bytes in reverse so that \x41\x42 ('AB')
872 * is equivalent to 0x4142. In other words, the bytes
873 * are in big-endian order. */
874 for (i = 0; i < n; i++) {
875 bytes[n - i - 1] = s[i + 1];
876 }
877 byten += n;
878 break;
eb254f26 879 } else if (s < t && (isDIGIT(*s) || *s == '-' || *s == '+')) {
569f27e5
JH
880 const char *u;
881 nantype =
882 grok_number_flags(s, (STRLEN)(t - s), &uv,
883 PERL_SCAN_TRAILING |
884 PERL_SCAN_ALLOW_UNDERSCORES);
885 /* Unfortunately grok_number_flags() doesn't
886 * tell how far we got and the ')' will always
887 * be "trailing", so we need to double-check
888 * whether we had something dubious. */
eb254f26
JH
889 u = s;
890 if ((*u == '-' || *u == '+')) {
891 u++;
892 }
893 for (; u < t; u++) {
569f27e5
JH
894 if (!isDIGIT(*u)) {
895 *flags |= IS_NUMBER_TRAILING;
896 break;
897 }
898 }
eb254f26
JH
899 if ((nantype & IS_NUMBER_NEG)) {
900 uv = (UV) (-uv);
901 }
569f27e5
JH
902 s = u;
903 } else {
eb254f26 904 bogus = TRUE;
569f27e5
JH
905 break;
906 }
907 /* XXX Doesn't do octal: nan("0123").
908 * Probably not a big loss. */
909
910 if (!(nantype & IS_NUMBER_IN_UV)) {
eb254f26 911 overflow = TRUE;
569f27e5
JH
912 break;
913 }
914
915 if (uv) {
eb254f26
JH
916 int bits = NV_NAN_BITS;
917 while (uv && byten < MAX_NV_BYTES && bits > 0) {
569f27e5
JH
918 bytes[byten++] = (U8) (uv & 0xFF);
919 uv >>= 8;
eb254f26 920 bits -= 8;
569f27e5
JH
921 }
922 }
eb254f26
JH
923 if (uv) {
924 overflow = TRUE;
925 }
569f27e5
JH
926 }
927
928 if (byten == 0) {
929 bytes[byten++] = 0;
930 }
931
4258cf90
JH
932 if (svp) {
933 if (bogus) {
934 sv_setpvf(svp, "NaN payload \"%s\" invalid",orig);
935 } else if (overflow) {
936 sv_setpvf(svp, "NaN payload \"%s\" overflowed %d bits",
937 orig, NV_NAN_BITS);
938 }
569f27e5
JH
939 }
940
941 if (s == send) {
942 *flags |= IS_NUMBER_TRAILING;
943 return s;
944 }
945
946 if (nvp) {
4258cf90 947 nan_payload_set(nvp, svp, bytes, byten, signaling);
569f27e5
JH
948 }
949
950 return s;
951}
952
953/*
954=for apidoc grok_nan
955
956Helper for grok_infnan().
957
958Parses the C99-style "nan(...)" strings, and sets the nvp accordingly.
959
960*sp points to the beginning of "nan", which can be also "qnan", "nanq",
961or "snan", "nans", and case is ignored.
962
963The "..." is parsed with grok_nan_payload().
964
965=cut
966*/
967const char *
4258cf90 968Perl_grok_nan(pTHX_ const char* s, const char* send, int *flags, NV* nvp, SV* svp)
569f27e5
JH
969{
970 bool signaling = FALSE;
971
972 PERL_ARGS_ASSERT_GROK_NAN;
973
974 if (isALPHA_FOLD_EQ(*s, 'S')) {
975 signaling = TRUE;
976 s++; if (s == send) return s;
977 } else if (isALPHA_FOLD_EQ(*s, 'Q')) {
978 s++; if (s == send) return s;
979 }
980
981 if (isALPHA_FOLD_EQ(*s, 'N')) {
982 s++; if (s == send || isALPHA_FOLD_NE(*s, 'A')) return s;
983 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return s;
984 s++;
985
986 *flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
987
988 /* NaN can be followed by various stuff (NaNQ, NaNS), while
989 * some legacy implementations have weird stuff like "NaN%"
990 * (no idea what that means). */
991 if (isALPHA_FOLD_EQ(*s, 's')) {
992 signaling = TRUE;
993 s++;
994 } else if (isALPHA_FOLD_EQ(*s, 'q')) {
995 s++;
996 }
997
998 if (*s == '(') {
4258cf90 999 const char *n = grok_nan_payload(s, send, signaling, flags, nvp, svp);
569f27e5
JH
1000 if (n == send) return NULL;
1001 s = n;
1002 if (*s != ')') {
1003 *flags |= IS_NUMBER_TRAILING;
1004 return s;
1005 }
1006 } else {
1007 if (nvp) {
1008 U8 bytes[1] = { 0 };
4258cf90 1009 nan_payload_set(nvp, svp, bytes, 1, signaling);
569f27e5
JH
1010 }
1011
1012 while (s < send && isSPACE(*s)) s++;
1013
1014 if (s < send && *s) {
1015 /* Note that we here implicitly accept (parse as
1016 * "nan", but with warnings) also any other weird
1017 * trailing stuff for "nan". In the above we just
1018 * check that if we got the C99-style "nan(...)",
1019 * the "..." looks sane. If in future we accept
1020 * more ways of specifying the nan payload (like
1021 * "nan123" or "nan0xabc"), the accepting would
1022 * happen around here. */
1023 *flags |= IS_NUMBER_TRAILING;
1024 }
1025 }
1026
1027 s = send;
1028 }
1029 else
1030 return NULL;
1031
1032 return s;
1033}
1034
1035/*
ff4eb398
JH
1036=for apidoc grok_infnan
1037
1038Helper for grok_number(), accepts various ways of spelling "infinity"
1039or "not a number", and returns one of the following flag combinations:
1040
1041 IS_NUMBER_INFINITE
1042 IS_NUMBER_NAN
1043 IS_NUMBER_INFINITE | IS_NUMBER_NEG
1044 IS_NUMBER_NAN | IS_NUMBER_NEG
1045 0
1046
62bdc035 1047possibly |-ed with IS_NUMBER_TRAILING.
b489e20f 1048
62bdc035
JH
1049If an infinity or a not-a-number is recognized, the *sp will point to
1050one byte past the end of the recognized string. If the recognition fails,
ff4eb398
JH
1051zero is returned, and the *sp will not move.
1052
1053=cut
1054*/
1055
1056int
4de1bcfe 1057Perl_grok_infnan(pTHX_ const char** sp, const char* send, NV* nvp)
ff4eb398
JH
1058{
1059 const char* s = *sp;
1060 int flags = 0;
62bdc035 1061 bool odh = FALSE; /* one-dot-hash: 1.#INF */
ff4eb398
JH
1062
1063 PERL_ARGS_ASSERT_GROK_INFNAN;
1064
4de1bcfe
JH
1065 /* XXX there are further legacy formats like HP-UX "++" for Inf
1066 * and "--" for -Inf. While we might be able to grok those in
1067 * string numification, having those in source code might open
1068 * up too much golfing: ++++;
1069 */
1070
8c12dc63
JH
1071 if (*s == '+') {
1072 s++; if (s == send) return 0;
1073 }
1074 else if (*s == '-') {
ff4eb398
JH
1075 flags |= IS_NUMBER_NEG; /* Yes, -NaN happens. Incorrect but happens. */
1076 s++; if (s == send) return 0;
1077 }
1078
1079 if (*s == '1') {
62bdc035
JH
1080 /* Visual C: 1.#SNAN, -1.#QNAN, 1#INF, 1.#IND (maybe also 1.#NAN)
1081 * Let's keep the dot optional. */
ff4eb398
JH
1082 s++; if (s == send) return 0;
1083 if (*s == '.') {
1084 s++; if (s == send) return 0;
1085 }
1086 if (*s == '#') {
1087 s++; if (s == send) return 0;
1088 } else
1089 return 0;
e855f543 1090 odh = TRUE;
ff4eb398
JH
1091 }
1092
305b8651 1093 if (isALPHA_FOLD_EQ(*s, 'I')) {
62bdc035
JH
1094 /* INF or IND (1.#IND is "indeterminate", a certain type of NAN) */
1095
305b8651 1096 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
ff4eb398 1097 s++; if (s == send) return 0;
305b8651 1098 if (isALPHA_FOLD_EQ(*s, 'F')) {
ff4eb398 1099 s++;
b8974fcb
JH
1100 if (s < send && (isALPHA_FOLD_EQ(*s, 'I'))) {
1101 int fail =
1102 flags | IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT | IS_NUMBER_TRAILING;
1103 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return fail;
1104 s++; if (s == send || isALPHA_FOLD_NE(*s, 'I')) return fail;
1105 s++; if (s == send || isALPHA_FOLD_NE(*s, 'T')) return fail;
1106 s++; if (s == send || isALPHA_FOLD_NE(*s, 'Y')) return fail;
3396ed30 1107 s++;
b8974fcb
JH
1108 } else if (odh) {
1109 while (*s == '0') { /* 1.#INF00 */
1110 s++;
1111 }
3396ed30 1112 }
b489e20f
JH
1113 while (s < send && isSPACE(*s))
1114 s++;
1115 if (s < send && *s) {
3396ed30 1116 flags |= IS_NUMBER_TRAILING;
fae4db12 1117 }
ff4eb398 1118 flags |= IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT;
4de1bcfe
JH
1119 if (nvp) {
1120 *nvp = (flags & IS_NUMBER_NEG) ? -NV_INF: NV_INF;
1121 }
ff4eb398 1122 }
e855f543 1123 else if (isALPHA_FOLD_EQ(*s, 'D') && odh) { /* 1.#IND */
ff4eb398
JH
1124 s++;
1125 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
4de1bcfe
JH
1126 if (nvp) {
1127 *nvp = NV_NAN;
1128 }
fae4db12
JH
1129 while (*s == '0') { /* 1.#IND00 */
1130 s++;
1131 }
1e9aa12f
JH
1132 if (*s) {
1133 flags |= IS_NUMBER_TRAILING;
1134 }
ff4eb398
JH
1135 } else
1136 return 0;
ff4eb398
JH
1137 }
1138 else {
62bdc035 1139 /* Maybe NAN of some sort */
4258cf90 1140 const char *n = grok_nan(s, send, &flags, nvp, NULL);
4de1bcfe
JH
1141 if (n == NULL) return 0;
1142 s = n;
ff4eb398
JH
1143 }
1144
b489e20f
JH
1145 while (s < send && isSPACE(*s))
1146 s++;
1147
a1fe7cea
JH
1148 *sp = s;
1149 return flags;
ff4eb398
JH
1150}
1151
13393a5e 1152/*
4de1bcfe 1153=for apidoc grok_number2_flags
13393a5e
JH
1154
1155Recognise (or not) a number. The type of the number is returned
1156(0 if unrecognised), otherwise it is a bit-ORed combination of
1157IS_NUMBER_IN_UV, IS_NUMBER_GREATER_THAN_UV_MAX, IS_NUMBER_NOT_INT,
1158IS_NUMBER_NEG, IS_NUMBER_INFINITY, IS_NUMBER_NAN (defined in perl.h).
1159
1160If the value of the number can fit in a UV, it is returned in the *valuep
1161IS_NUMBER_IN_UV will be set to indicate that *valuep is valid, IS_NUMBER_IN_UV
1162will never be set unless *valuep is valid, but *valuep may have been assigned
1163to during processing even though IS_NUMBER_IN_UV is not set on return.
1164If valuep is NULL, IS_NUMBER_IN_UV will be set for the same cases as when
1165valuep is non-NULL, but no actual assignment (or SEGV) will occur.
1166
4de1bcfe
JH
1167The nvp is used to directly set the value for infinities (Inf) and
1168not-a-numbers (NaN).
1169
13393a5e
JH
1170IS_NUMBER_NOT_INT will be set with IS_NUMBER_IN_UV if trailing decimals were
1171seen (in which case *valuep gives the true value truncated to an integer), and
1172IS_NUMBER_NEG if the number is negative (in which case *valuep holds the
1173absolute value). IS_NUMBER_IN_UV is not set if e notation was used or the
1174number is larger than a UV.
1175
1176C<flags> allows only C<PERL_SCAN_TRAILING>, which allows for trailing
1177non-numeric text on an otherwise successful I<grok>, setting
1178C<IS_NUMBER_TRAILING> on the result.
1179
4de1bcfe
JH
1180=for apidoc grok_number_flags
1181
1182Identical to grok_number2_flags() with nvp and flags set to zero.
1183
13393a5e
JH
1184=for apidoc grok_number
1185
1186Identical to grok_number_flags() with flags set to zero.
1187
1188=cut
1189 */
1190int
1191Perl_grok_number(pTHX_ const char *pv, STRLEN len, UV *valuep)
1192{
1193 PERL_ARGS_ASSERT_GROK_NUMBER;
1194
1195 return grok_number_flags(pv, len, valuep, 0);
1196}
1197
4de1bcfe
JH
1198int
1199Perl_grok_number_flags(pTHX_ const char *pv, STRLEN len, UV *valuep, U32 flags)
1200{
1201 PERL_ARGS_ASSERT_GROK_NUMBER_FLAGS;
1202
1203 return grok_number2_flags(pv, len, valuep, NULL, flags);
1204}
1205
945b524a
JH
1206static const UV uv_max_div_10 = UV_MAX / 10;
1207static const U8 uv_max_mod_10 = UV_MAX % 10;
1208
3f7602fa 1209int
4de1bcfe 1210Perl_grok_number2_flags(pTHX_ const char *pv, STRLEN len, UV *valuep, NV *nvp, U32 flags)
3f7602fa 1211{
60939fb8 1212 const char *s = pv;
c4420975 1213 const char * const send = pv + len;
ae776a2c 1214 const char *d;
60939fb8 1215 int numtype = 0;
60939fb8 1216
4de1bcfe 1217 PERL_ARGS_ASSERT_GROK_NUMBER2_FLAGS;
7918f24d 1218
60939fb8
NC
1219 while (s < send && isSPACE(*s))
1220 s++;
1221 if (s == send) {
1222 return 0;
1223 } else if (*s == '-') {
1224 s++;
1225 numtype = IS_NUMBER_NEG;
1226 }
1227 else if (*s == '+')
aa42a541 1228 s++;
60939fb8
NC
1229
1230 if (s == send)
1231 return 0;
1232
ae776a2c 1233 /* The first digit (after optional sign): note that might
8c12dc63 1234 * also point to "infinity" or "nan", or "1.#INF". */
ae776a2c
JH
1235 d = s;
1236
8c12dc63 1237 /* next must be digit or the radix separator or beginning of infinity/nan */
60939fb8
NC
1238 if (isDIGIT(*s)) {
1239 /* UVs are at least 32 bits, so the first 9 decimal digits cannot
1240 overflow. */
1241 UV value = *s - '0';
1242 /* This construction seems to be more optimiser friendly.
1243 (without it gcc does the isDIGIT test and the *s - '0' separately)
1244 With it gcc on arm is managing 6 instructions (6 cycles) per digit.
1245 In theory the optimiser could deduce how far to unroll the loop
1246 before checking for overflow. */
58bb9ec3
NC
1247 if (++s < send) {
1248 int digit = *s - '0';
60939fb8
NC
1249 if (digit >= 0 && digit <= 9) {
1250 value = value * 10 + digit;
58bb9ec3
NC
1251 if (++s < send) {
1252 digit = *s - '0';
60939fb8
NC
1253 if (digit >= 0 && digit <= 9) {
1254 value = value * 10 + digit;
58bb9ec3
NC
1255 if (++s < send) {
1256 digit = *s - '0';
60939fb8
NC
1257 if (digit >= 0 && digit <= 9) {
1258 value = value * 10 + digit;
58bb9ec3
NC
1259 if (++s < send) {
1260 digit = *s - '0';
60939fb8
NC
1261 if (digit >= 0 && digit <= 9) {
1262 value = value * 10 + digit;
58bb9ec3
NC
1263 if (++s < send) {
1264 digit = *s - '0';
60939fb8
NC
1265 if (digit >= 0 && digit <= 9) {
1266 value = value * 10 + digit;
58bb9ec3
NC
1267 if (++s < send) {
1268 digit = *s - '0';
60939fb8
NC
1269 if (digit >= 0 && digit <= 9) {
1270 value = value * 10 + digit;
58bb9ec3
NC
1271 if (++s < send) {
1272 digit = *s - '0';
60939fb8
NC
1273 if (digit >= 0 && digit <= 9) {
1274 value = value * 10 + digit;
58bb9ec3
NC
1275 if (++s < send) {
1276 digit = *s - '0';
60939fb8
NC
1277 if (digit >= 0 && digit <= 9) {
1278 value = value * 10 + digit;
58bb9ec3 1279 if (++s < send) {
60939fb8
NC
1280 /* Now got 9 digits, so need to check
1281 each time for overflow. */
58bb9ec3 1282 digit = *s - '0';
60939fb8 1283 while (digit >= 0 && digit <= 9
945b524a
JH
1284 && (value < uv_max_div_10
1285 || (value == uv_max_div_10
1286 && digit <= uv_max_mod_10))) {
60939fb8 1287 value = value * 10 + digit;
58bb9ec3
NC
1288 if (++s < send)
1289 digit = *s - '0';
60939fb8
NC
1290 else
1291 break;
1292 }
1293 if (digit >= 0 && digit <= 9
51bd16da 1294 && (s < send)) {
60939fb8
NC
1295 /* value overflowed.
1296 skip the remaining digits, don't
1297 worry about setting *valuep. */
1298 do {
1299 s++;
1300 } while (s < send && isDIGIT(*s));
1301 numtype |=
1302 IS_NUMBER_GREATER_THAN_UV_MAX;
1303 goto skip_value;
1304 }
1305 }
1306 }
98994639 1307 }
60939fb8
NC
1308 }
1309 }
1310 }
1311 }
1312 }
1313 }
1314 }
1315 }
1316 }
1317 }
1318 }
98994639 1319 }
60939fb8 1320 }
98994639 1321 }
60939fb8
NC
1322 numtype |= IS_NUMBER_IN_UV;
1323 if (valuep)
1324 *valuep = value;
1325
1326 skip_value:
1327 if (GROK_NUMERIC_RADIX(&s, send)) {
1328 numtype |= IS_NUMBER_NOT_INT;
1329 while (s < send && isDIGIT(*s)) /* optional digits after the radix */
1330 s++;
98994639 1331 }
60939fb8
NC
1332 }
1333 else if (GROK_NUMERIC_RADIX(&s, send)) {
1334 numtype |= IS_NUMBER_NOT_INT | IS_NUMBER_IN_UV; /* valuep assigned below */
1335 /* no digits before the radix means we need digits after it */
1336 if (s < send && isDIGIT(*s)) {
1337 do {
1338 s++;
1339 } while (s < send && isDIGIT(*s));
1340 if (valuep) {
1341 /* integer approximation is valid - it's 0. */
1342 *valuep = 0;
1343 }
98994639 1344 }
60939fb8 1345 else
ae776a2c 1346 return 0;
ff4eb398 1347 }
60939fb8 1348
926f5fc6 1349 if (s > d && s < send) {
60939fb8 1350 /* we can have an optional exponent part */
305b8651 1351 if (isALPHA_FOLD_EQ(*s, 'e')) {
60939fb8
NC
1352 s++;
1353 if (s < send && (*s == '-' || *s == '+'))
1354 s++;
1355 if (s < send && isDIGIT(*s)) {
1356 do {
1357 s++;
1358 } while (s < send && isDIGIT(*s));
1359 }
3f7602fa
TC
1360 else if (flags & PERL_SCAN_TRAILING)
1361 return numtype | IS_NUMBER_TRAILING;
60939fb8 1362 else
3f7602fa
TC
1363 return 0;
1364
1365 /* The only flag we keep is sign. Blow away any "it's UV" */
1366 numtype &= IS_NUMBER_NEG;
1367 numtype |= IS_NUMBER_NOT_INT;
60939fb8
NC
1368 }
1369 }
1370 while (s < send && isSPACE(*s))
1371 s++;
1372 if (s >= send)
aa8b85de 1373 return numtype;
60939fb8
NC
1374 if (len == 10 && memEQ(pv, "0 but true", 10)) {
1375 if (valuep)
1376 *valuep = 0;
1377 return IS_NUMBER_IN_UV;
1378 }
8c12dc63
JH
1379 /* We could be e.g. at "Inf" or "NaN", or at the "#" of "1.#INF". */
1380 if ((s + 2 < send) && strchr("inqs#", toFOLD(*s))) {
1381 /* Really detect inf/nan. Start at d, not s, since the above
1382 * code might have already consumed the "1." or "1". */
4de1bcfe
JH
1383 NV nanv;
1384 int infnan = Perl_grok_infnan(aTHX_ &d, send, &nanv);
8c12dc63 1385 if ((infnan & IS_NUMBER_INFINITY)) {
4de1bcfe
JH
1386 if (nvp) {
1387 *nvp = (numtype & IS_NUMBER_NEG) ? -NV_INF : NV_INF;
1388 }
8c12dc63
JH
1389 return (numtype | infnan); /* Keep sign for infinity. */
1390 }
1391 else if ((infnan & IS_NUMBER_NAN)) {
4de1bcfe
JH
1392 if (nvp) {
1393 *nvp = nanv;
1394 }
8c12dc63
JH
1395 return (numtype | infnan) & ~IS_NUMBER_NEG; /* Clear sign for nan. */
1396 }
1397 }
3f7602fa
TC
1398 else if (flags & PERL_SCAN_TRAILING) {
1399 return numtype | IS_NUMBER_TRAILING;
1400 }
1401
60939fb8 1402 return 0;
98994639
HS
1403}
1404
6313e544 1405/*
d62b8c6a 1406=for apidoc grok_atou
6313e544 1407
d62b8c6a 1408grok_atou is a safer replacement for atoi and strtol.
6313e544 1409
d62b8c6a
JH
1410grok_atou parses a C-style zero-byte terminated string, looking for
1411a decimal unsigned integer.
338aa8b0 1412
d62b8c6a
JH
1413Returns the unsigned integer, if a valid value can be parsed
1414from the beginning of the string.
f4379102 1415
d62b8c6a 1416Accepts only the decimal digits '0'..'9'.
6313e544 1417
d62b8c6a
JH
1418As opposed to atoi or strtol, grok_atou does NOT allow optional
1419leading whitespace, or negative inputs. If such features are
1420required, the calling code needs to explicitly implement those.
6313e544 1421
d62b8c6a 1422If a valid value cannot be parsed, returns either zero (if non-digits
75feedba 1423are met before any digits) or UV_MAX (if the value overflows).
6313e544 1424
d62b8c6a
JH
1425Note that extraneous leading zeros also count as an overflow
1426(meaning that only "0" is the zero).
338aa8b0 1427
d62b8c6a 1428On failure, the *endptr is also set to NULL, unless endptr is NULL.
338aa8b0
JH
1429
1430Trailing non-digit bytes are allowed if the endptr is non-NULL.
6313e544
JH
1431On return the *endptr will contain the pointer to the first non-digit byte.
1432
6313e544 1433If the endptr is NULL, the first non-digit byte MUST be
f4379102 1434the zero byte terminating the pv, or zero will be returned.
6313e544 1435
d62b8c6a
JH
1436Background: atoi has severe problems with illegal inputs, it cannot be
1437used for incremental parsing, and therefore should be avoided
1438atoi and strtol are also affected by locale settings, which can also be
1439seen as a bug (global state controlled by user environment).
1440
6313e544
JH
1441=cut
1442*/
1443
75feedba 1444UV
6313e544
JH
1445Perl_grok_atou(const char *pv, const char** endptr)
1446{
1447 const char* s = pv;
1448 const char** eptr;
1449 const char* end2; /* Used in case endptr is NULL. */
75feedba 1450 UV val = 0; /* The return value. */
6313e544
JH
1451
1452 PERL_ARGS_ASSERT_GROK_ATOU;
1453
1454 eptr = endptr ? endptr : &end2;
75feedba
JH
1455 if (isDIGIT(*s)) {
1456 /* Single-digit inputs are quite common. */
6313e544 1457 val = *s++ - '0';
75feedba
JH
1458 if (isDIGIT(*s)) {
1459 /* Extra leading zeros cause overflow. */
1460 if (val == 0) {
1461 *eptr = NULL;
1462 return UV_MAX;
1463 }
1464 while (isDIGIT(*s)) {
1465 /* This could be unrolled like in grok_number(), but
1466 * the expected uses of this are not speed-needy, and
1467 * unlikely to need full 64-bitness. */
1468 U8 digit = *s++ - '0';
945b524a
JH
1469 if (val < uv_max_div_10 ||
1470 (val == uv_max_div_10 && digit <= uv_max_mod_10)) {
75feedba
JH
1471 val = val * 10 + digit;
1472 } else {
6313e544 1473 *eptr = NULL;
75feedba 1474 return UV_MAX;
6313e544 1475 }
6313e544
JH
1476 }
1477 }
75feedba
JH
1478 }
1479 if (s == pv) {
1480 *eptr = NULL; /* If no progress, failed to parse anything. */
1481 return 0;
6313e544
JH
1482 }
1483 if (endptr == NULL && *s) {
1484 return 0; /* If endptr is NULL, no trailing non-digits allowed. */
1485 }
1486 *eptr = s;
1487 return val;
1488}
1489
a4eca1d4 1490#ifndef USE_QUADMATH
4801ca72 1491STATIC NV
98994639
HS
1492S_mulexp10(NV value, I32 exponent)
1493{
1494 NV result = 1.0;
1495 NV power = 10.0;
1496 bool negative = 0;
1497 I32 bit;
1498
1499 if (exponent == 0)
1500 return value;
659c4b96
DM
1501 if (value == 0)
1502 return (NV)0;
87032ba1 1503
24866caa 1504 /* On OpenVMS VAX we by default use the D_FLOAT double format,
67597c89 1505 * and that format does not have *easy* capabilities [1] for
24866caa
CB
1506 * overflowing doubles 'silently' as IEEE fp does. We also need
1507 * to support G_FLOAT on both VAX and Alpha, and though the exponent
1508 * range is much larger than D_FLOAT it still doesn't do silent
1509 * overflow. Therefore we need to detect early whether we would
1510 * overflow (this is the behaviour of the native string-to-float
1511 * conversion routines, and therefore of native applications, too).
67597c89 1512 *
24866caa
CB
1513 * [1] Trying to establish a condition handler to trap floating point
1514 * exceptions is not a good idea. */
87032ba1
JH
1515
1516 /* In UNICOS and in certain Cray models (such as T90) there is no
1517 * IEEE fp, and no way at all from C to catch fp overflows gracefully.
1518 * There is something you can do if you are willing to use some
1519 * inline assembler: the instruction is called DFI-- but that will
1520 * disable *all* floating point interrupts, a little bit too large
1521 * a hammer. Therefore we need to catch potential overflows before
1522 * it's too late. */
353813d9 1523
85bba25f 1524#if ((defined(VMS) && !defined(_IEEE_FP)) || defined(_UNICOS)) && defined(NV_MAX_10_EXP)
353813d9 1525 STMT_START {
c4420975 1526 const NV exp_v = log10(value);
353813d9
HS
1527 if (exponent >= NV_MAX_10_EXP || exponent + exp_v >= NV_MAX_10_EXP)
1528 return NV_MAX;
1529 if (exponent < 0) {
1530 if (-(exponent + exp_v) >= NV_MAX_10_EXP)
1531 return 0.0;
1532 while (-exponent >= NV_MAX_10_EXP) {
1533 /* combination does not overflow, but 10^(-exponent) does */
1534 value /= 10;
1535 ++exponent;
1536 }
1537 }
1538 } STMT_END;
87032ba1
JH
1539#endif
1540
353813d9
HS
1541 if (exponent < 0) {
1542 negative = 1;
1543 exponent = -exponent;
b27804d8
DM
1544#ifdef NV_MAX_10_EXP
1545 /* for something like 1234 x 10^-309, the action of calculating
1546 * the intermediate value 10^309 then returning 1234 / (10^309)
1547 * will fail, since 10^309 becomes infinity. In this case try to
1548 * refactor it as 123 / (10^308) etc.
1549 */
1550 while (value && exponent > NV_MAX_10_EXP) {
1551 exponent--;
1552 value /= 10;
1553 }
48853916
JH
1554 if (value == 0.0)
1555 return value;
b27804d8 1556#endif
353813d9 1557 }
c62e754c
JH
1558#if defined(__osf__)
1559 /* Even with cc -ieee + ieee_set_fp_control(IEEE_TRAP_ENABLE_INV)
1560 * Tru64 fp behavior on inf/nan is somewhat broken. Another way
1561 * to do this would be ieee_set_fp_control(IEEE_TRAP_ENABLE_OVF)
1562 * but that breaks another set of infnan.t tests. */
1563# define FP_OVERFLOWS_TO_ZERO
1564#endif
98994639
HS
1565 for (bit = 1; exponent; bit <<= 1) {
1566 if (exponent & bit) {
1567 exponent ^= bit;
1568 result *= power;
c62e754c
JH
1569#ifdef FP_OVERFLOWS_TO_ZERO
1570 if (result == 0)
1571 return value < 0 ? -NV_INF : NV_INF;
1572#endif
236f0012
CB
1573 /* Floating point exceptions are supposed to be turned off,
1574 * but if we're obviously done, don't risk another iteration.
1575 */
1576 if (exponent == 0) break;
98994639
HS
1577 }
1578 power *= power;
1579 }
1580 return negative ? value / result : value * result;
1581}
a4eca1d4 1582#endif /* #ifndef USE_QUADMATH */
98994639
HS
1583
1584NV
1585Perl_my_atof(pTHX_ const char* s)
1586{
1587 NV x = 0.0;
a4eca1d4
JH
1588#ifdef USE_QUADMATH
1589 Perl_my_atof2(aTHX_ s, &x);
1590 return x;
1591#else
1592# ifdef USE_LOCALE_NUMERIC
7918f24d
NC
1593 PERL_ARGS_ASSERT_MY_ATOF;
1594
a2287a13
KW
1595 {
1596 DECLARE_STORE_LC_NUMERIC_SET_TO_NEEDED();
d6ded950 1597 if (PL_numeric_radix_sv && IN_LC(LC_NUMERIC)) {
e4850248
KW
1598 const char *standard = NULL, *local = NULL;
1599 bool use_standard_radix;
98994639 1600
e4850248
KW
1601 /* Look through the string for the first thing that looks like a
1602 * decimal point: either the value in the current locale or the
1603 * standard fallback of '.'. The one which appears earliest in the
1604 * input string is the one that we should have atof look for. Note
1605 * that we have to determine this beforehand because on some
1606 * systems, Perl_atof2 is just a wrapper around the system's atof.
1607 * */
1608 standard = strchr(s, '.');
1609 local = strstr(s, SvPV_nolen(PL_numeric_radix_sv));
78787052 1610
e4850248 1611 use_standard_radix = standard && (!local || standard < local);
78787052 1612
e4850248
KW
1613 if (use_standard_radix)
1614 SET_NUMERIC_STANDARD();
78787052 1615
e4850248 1616 Perl_atof2(s, x);
78787052 1617
e4850248
KW
1618 if (use_standard_radix)
1619 SET_NUMERIC_LOCAL();
1620 }
1621 else
1622 Perl_atof2(s, x);
a2287a13
KW
1623 RESTORE_LC_NUMERIC();
1624 }
a4eca1d4 1625# else
a36244b7 1626 Perl_atof2(s, x);
a4eca1d4 1627# endif
98994639
HS
1628#endif
1629 return x;
1630}
1631
3c81f0b3
DD
1632
1633#ifdef USING_MSVC6
1634# pragma warning(push)
1635# pragma warning(disable:4756;disable:4056)
1636#endif
829757a4 1637static char*
5563f457 1638S_my_atof_infnan(pTHX_ const char* s, bool negative, const char* send, NV* value)
829757a4
JH
1639{
1640 const char *p0 = negative ? s - 1 : s;
1641 const char *p = p0;
4de1bcfe 1642 int infnan = grok_infnan(&p, send, value);
829757a4
JH
1643 if (infnan && p != p0) {
1644 /* If we can generate inf/nan directly, let's do so. */
1645#ifdef NV_INF
1646 if ((infnan & IS_NUMBER_INFINITY)) {
4de1bcfe 1647 /* grok_infnan() already set the value. */
829757a4
JH
1648 return (char*)p;
1649 }
1650#endif
1651#ifdef NV_NAN
1652 if ((infnan & IS_NUMBER_NAN)) {
4de1bcfe 1653 /* grok_infnan() already set the value. */
829757a4
JH
1654 return (char*)p;
1655 }
1656#endif
1657#ifdef Perl_strtod
68611e6f 1658 /* If still here, we didn't have either NV_INF or NV_NAN,
829757a4
JH
1659 * and can try falling back to native strtod/strtold.
1660 *
68611e6f
JH
1661 * (Though, are our NV_INF or NV_NAN ever not defined?)
1662 *
829757a4
JH
1663 * The native interface might not recognize all the possible
1664 * inf/nan strings Perl recognizes. What we can try
1665 * is to try faking the input. We will try inf/-inf/nan
1666 * as the most promising/portable input. */
1667 {
1668 const char* fake = NULL;
1669 char* endp;
1670 NV nv;
1671 if ((infnan & IS_NUMBER_INFINITY)) {
1672 fake = ((infnan & IS_NUMBER_NEG)) ? "-inf" : "inf";
1673 }
1674 else if ((infnan & IS_NUMBER_NAN)) {
1675 fake = "nan";
1676 }
1677 assert(fake);
1678 nv = Perl_strtod(fake, &endp);
1679 if (fake != endp) {
1680 if ((infnan & IS_NUMBER_INFINITY)) {
1681#ifdef Perl_isinf
1682 if (Perl_isinf(nv))
1683 *value = nv;
1684#else
1685 /* last resort, may generate SIGFPE */
1686 *value = Perl_exp((NV)1e9);
1687 if ((infnan & IS_NUMBER_NEG))
1688 *value = -*value;
1689#endif
1690 return (char*)p; /* p, not endp */
1691 }
1692 else if ((infnan & IS_NUMBER_NAN)) {
1693#ifdef Perl_isnan
1694 if (Perl_isnan(nv))
1695 *value = nv;
1696#else
1697 /* last resort, may generate SIGFPE */
1698 *value = Perl_log((NV)-1.0);
1699#endif
1700 return (char*)p; /* p, not endp */
1701 }
1702 }
1703 }
1704#endif /* #ifdef Perl_strtod */
1705 }
1706 return NULL;
1707}
3c81f0b3
DD
1708#ifdef USING_MSVC6
1709# pragma warning(pop)
1710#endif
829757a4 1711
98994639
HS
1712char*
1713Perl_my_atof2(pTHX_ const char* orig, NV* value)
1714{
e1ec3a88 1715 const char* s = orig;
a4eca1d4
JH
1716 NV result[3] = {0.0, 0.0, 0.0};
1717#if defined(USE_PERL_ATOF) || defined(USE_QUADMATH)
ae776a2c 1718 const char* send = s + strlen(orig); /* one past the last */
a4eca1d4
JH
1719 bool negative = 0;
1720#endif
1721#if defined(USE_PERL_ATOF) && !defined(USE_QUADMATH)
1722 UV accumulator[2] = {0,0}; /* before/after dp */
8194bf88 1723 bool seen_digit = 0;
20f6aaab
AS
1724 I32 exp_adjust[2] = {0,0};
1725 I32 exp_acc[2] = {-1, -1};
1726 /* the current exponent adjust for the accumulators */
98994639 1727 I32 exponent = 0;
8194bf88 1728 I32 seen_dp = 0;
20f6aaab
AS
1729 I32 digit = 0;
1730 I32 old_digit = 0;
8194bf88 1731 I32 sig_digits = 0; /* noof significant digits seen so far */
a4eca1d4 1732#endif
8194bf88 1733
a4eca1d4 1734#if defined(USE_PERL_ATOF) || defined(USE_QUADMATH)
7918f24d
NC
1735 PERL_ARGS_ASSERT_MY_ATOF2;
1736
a4eca1d4
JH
1737 /* leading whitespace */
1738 while (isSPACE(*s))
1739 ++s;
1740
1741 /* sign */
1742 switch (*s) {
1743 case '-':
1744 negative = 1;
1745 /* FALLTHROUGH */
1746 case '+':
1747 ++s;
1748 }
1749#endif
1750
1751#ifdef USE_QUADMATH
1752 {
1753 char* endp;
1754 if ((endp = S_my_atof_infnan(s, negative, send, value)))
1755 return endp;
1756 result[2] = strtoflt128(s, &endp);
1757 if (s != endp) {
1758 *value = negative ? -result[2] : result[2];
1759 return endp;
1760 }
1761 return NULL;
1762 }
1763#elif defined(USE_PERL_ATOF)
1764
8194bf88
DM
1765/* There is no point in processing more significant digits
1766 * than the NV can hold. Note that NV_DIG is a lower-bound value,
1767 * while we need an upper-bound value. We add 2 to account for this;
1768 * since it will have been conservative on both the first and last digit.
1769 * For example a 32-bit mantissa with an exponent of 4 would have
1770 * exact values in the set
1771 * 4
1772 * 8
1773 * ..
1774 * 17179869172
1775 * 17179869176
1776 * 17179869180
1777 *
1778 * where for the purposes of calculating NV_DIG we would have to discount
1779 * both the first and last digit, since neither can hold all values from
1780 * 0..9; but for calculating the value we must examine those two digits.
1781 */
ffa277e5
EAS
1782#ifdef MAX_SIG_DIG_PLUS
1783 /* It is not necessarily the case that adding 2 to NV_DIG gets all the
1784 possible digits in a NV, especially if NVs are not IEEE compliant
1785 (e.g., long doubles on IRIX) - Allen <allens@cpan.org> */
1786# define MAX_SIG_DIGITS (NV_DIG+MAX_SIG_DIG_PLUS)
1787#else
1788# define MAX_SIG_DIGITS (NV_DIG+2)
1789#endif
8194bf88
DM
1790
1791/* the max number we can accumulate in a UV, and still safely do 10*N+9 */
1792#define MAX_ACCUMULATE ( (UV) ((UV_MAX - 9)/10))
98994639 1793
ae776a2c 1794 {
829757a4 1795 const char* endp;
5563f457 1796 if ((endp = S_my_atof_infnan(aTHX_ s, negative, send, value)))
829757a4 1797 return (char*)endp;
ae776a2c 1798 }
2b54f59f 1799
8194bf88
DM
1800 /* we accumulate digits into an integer; when this becomes too
1801 * large, we add the total to NV and start again */
98994639 1802
8194bf88
DM
1803 while (1) {
1804 if (isDIGIT(*s)) {
1805 seen_digit = 1;
20f6aaab 1806 old_digit = digit;
8194bf88 1807 digit = *s++ - '0';
20f6aaab
AS
1808 if (seen_dp)
1809 exp_adjust[1]++;
98994639 1810
8194bf88
DM
1811 /* don't start counting until we see the first significant
1812 * digit, eg the 5 in 0.00005... */
1813 if (!sig_digits && digit == 0)
1814 continue;
1815
1816 if (++sig_digits > MAX_SIG_DIGITS) {
98994639 1817 /* limits of precision reached */
20f6aaab
AS
1818 if (digit > 5) {
1819 ++accumulator[seen_dp];
1820 } else if (digit == 5) {
1821 if (old_digit % 2) { /* round to even - Allen */
1822 ++accumulator[seen_dp];
1823 }
1824 }
1825 if (seen_dp) {
1826 exp_adjust[1]--;
1827 } else {
1828 exp_adjust[0]++;
1829 }
8194bf88 1830 /* skip remaining digits */
98994639 1831 while (isDIGIT(*s)) {
98994639 1832 ++s;
20f6aaab
AS
1833 if (! seen_dp) {
1834 exp_adjust[0]++;
1835 }
98994639
HS
1836 }
1837 /* warn of loss of precision? */
98994639 1838 }
8194bf88 1839 else {
20f6aaab 1840 if (accumulator[seen_dp] > MAX_ACCUMULATE) {
8194bf88 1841 /* add accumulator to result and start again */
20f6aaab
AS
1842 result[seen_dp] = S_mulexp10(result[seen_dp],
1843 exp_acc[seen_dp])
1844 + (NV)accumulator[seen_dp];
1845 accumulator[seen_dp] = 0;
1846 exp_acc[seen_dp] = 0;
98994639 1847 }
20f6aaab
AS
1848 accumulator[seen_dp] = accumulator[seen_dp] * 10 + digit;
1849 ++exp_acc[seen_dp];
98994639 1850 }
8194bf88 1851 }
e1ec3a88 1852 else if (!seen_dp && GROK_NUMERIC_RADIX(&s, send)) {
8194bf88 1853 seen_dp = 1;
20f6aaab 1854 if (sig_digits > MAX_SIG_DIGITS) {
c86f7df5 1855 do {
20f6aaab 1856 ++s;
c86f7df5 1857 } while (isDIGIT(*s));
20f6aaab
AS
1858 break;
1859 }
8194bf88
DM
1860 }
1861 else {
1862 break;
98994639
HS
1863 }
1864 }
1865
20f6aaab
AS
1866 result[0] = S_mulexp10(result[0], exp_acc[0]) + (NV)accumulator[0];
1867 if (seen_dp) {
1868 result[1] = S_mulexp10(result[1], exp_acc[1]) + (NV)accumulator[1];
1869 }
98994639 1870
305b8651 1871 if (seen_digit && (isALPHA_FOLD_EQ(*s, 'e'))) {
98994639
HS
1872 bool expnegative = 0;
1873
1874 ++s;
1875 switch (*s) {
1876 case '-':
1877 expnegative = 1;
924ba076 1878 /* FALLTHROUGH */
98994639
HS
1879 case '+':
1880 ++s;
1881 }
1882 while (isDIGIT(*s))
1883 exponent = exponent * 10 + (*s++ - '0');
1884 if (expnegative)
1885 exponent = -exponent;
1886 }
1887
20f6aaab
AS
1888
1889
98994639 1890 /* now apply the exponent */
20f6aaab
AS
1891
1892 if (seen_dp) {
1893 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0])
1894 + S_mulexp10(result[1],exponent-exp_adjust[1]);
1895 } else {
1896 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0]);
1897 }
98994639
HS
1898
1899 /* now apply the sign */
1900 if (negative)
20f6aaab 1901 result[2] = -result[2];
a36244b7 1902#endif /* USE_PERL_ATOF */
20f6aaab 1903 *value = result[2];
73d840c0 1904 return (char *)s;
98994639
HS
1905}
1906
5d34af89 1907/*
3d9d9213 1908=for apidoc isinfnan
5d34af89
JH
1909
1910Perl_isinfnan() is utility function that returns true if the NV
1911argument is either an infinity or a NaN, false otherwise. To test
1912in more detail, use Perl_isinf() and Perl_isnan().
1913
68611e6f
JH
1914This is also the logical inverse of Perl_isfinite().
1915
5d34af89
JH
1916=cut
1917*/
1cd88304
JH
1918bool
1919Perl_isinfnan(NV nv)
1920{
1921#ifdef Perl_isinf
1922 if (Perl_isinf(nv))
1923 return TRUE;
1924#endif
1925#ifdef Perl_isnan
1926 if (Perl_isnan(nv))
1927 return TRUE;
1928#endif
1929 return FALSE;
1930}
1931
354b74ae
FC
1932/*
1933=for apidoc
1934
1935Checks whether the argument would be either an infinity or NaN when used
1936as a number, but is careful not to trigger non-numeric or uninitialized
1937warnings. it assumes the caller has done SvGETMAGIC(sv) already.
1938
1939=cut
1940*/
1941
1942bool
1943Perl_isinfnansv(pTHX_ SV *sv)
1944{
1945 PERL_ARGS_ASSERT_ISINFNANSV;
1946 if (!SvOK(sv))
1947 return FALSE;
1948 if (SvNOKp(sv))
1949 return Perl_isinfnan(SvNVX(sv));
1950 if (SvIOKp(sv))
1951 return FALSE;
1952 {
1953 STRLEN len;
1954 const char *s = SvPV_nomg_const(sv, len);
4de1bcfe 1955 return cBOOL(grok_infnan(&s, s+len, NULL));
354b74ae
FC
1956 }
1957}
1958
d67dac15 1959#ifndef HAS_MODFL
68611e6f
JH
1960/* C99 has truncl, pre-C99 Solaris had aintl. We can use either with
1961 * copysignl to emulate modfl, which is in some platforms missing or
1962 * broken. */
d67dac15
JH
1963# if defined(HAS_TRUNCL) && defined(HAS_COPYSIGNL)
1964long double
1965Perl_my_modfl(long double x, long double *ip)
1966{
68611e6f
JH
1967 *ip = truncl(x);
1968 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
d67dac15
JH
1969}
1970# elif defined(HAS_AINTL) && defined(HAS_COPYSIGNL)
55954f19
JH
1971long double
1972Perl_my_modfl(long double x, long double *ip)
1973{
68611e6f
JH
1974 *ip = aintl(x);
1975 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
55954f19 1976}
d67dac15 1977# endif
55954f19
JH
1978#endif
1979
7b9b7dff 1980/* Similarly, with ilogbl and scalbnl we can emulate frexpl. */
55954f19
JH
1981#if ! defined(HAS_FREXPL) && defined(HAS_ILOGBL) && defined(HAS_SCALBNL)
1982long double
1983Perl_my_frexpl(long double x, int *e) {
68611e6f
JH
1984 *e = x == 0.0L ? 0 : ilogbl(x) + 1;
1985 return (scalbnl(x, -*e));
55954f19
JH
1986}
1987#endif
66610fdd
RGS
1988
1989/*
ed140128
AD
1990=for apidoc Perl_signbit
1991
1992Return a non-zero integer if the sign bit on an NV is set, and 0 if
1993it is not.
1994
1995If Configure detects this system has a signbit() that will work with
1996our NVs, then we just use it via the #define in perl.h. Otherwise,
8b7fad81
JH
1997fall back on this implementation. The main use of this function
1998is catching -0.0.
ed140128
AD
1999
2000Configure notes: This function is called 'Perl_signbit' instead of a
2001plain 'signbit' because it is easy to imagine a system having a signbit()
2002function or macro that doesn't happen to work with our particular choice
2003of NVs. We shouldn't just re-#define signbit as Perl_signbit and expect
2004the standard system headers to be happy. Also, this is a no-context
2005function (no pTHX_) because Perl_signbit() is usually re-#defined in
2006perl.h as a simple macro call to the system's signbit().
2007Users should just always call Perl_signbit().
2008
2009=cut
2010*/
2011#if !defined(HAS_SIGNBIT)
2012int
2013Perl_signbit(NV x) {
8b7fad81
JH
2014# ifdef Perl_fp_class_nzero
2015 if (x == 0)
2016 return Perl_fp_class_nzero(x);
8b7fad81 2017# endif
3585840c 2018 return (x < 0.0) ? 1 : 0;
ed140128
AD
2019}
2020#endif
2021
2022/*
66610fdd
RGS
2023 * Local variables:
2024 * c-indentation-style: bsd
2025 * c-basic-offset: 4
14d04a33 2026 * indent-tabs-mode: nil
66610fdd
RGS
2027 * End:
2028 *
14d04a33 2029 * ex: set ts=8 sts=4 sw=4 et:
37442d52 2030 */