This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
APItest: Add U8* typemap, and use it
[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;
071db91b 42 return ((U32) f) | (1 + (U32_MAX >> 1));
98994639
HS
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;
071db91b 60 return (I32)(((U32) f) | (1 + (U32_MAX >> 1)));
98994639
HS
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;
071db91b 79 return (IV)(((UV) f) | (1 + (UV_MAX >> 1)));
98994639
HS
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;
071db91b 97 return ((UV) f) | (1 + (UV_MAX >> 1));
98994639
HS
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
2d7f6611 110On entry C<start> and C<*len> give the string to scan, C<*flags> gives
796b6530 111conversion flags, and C<result> should be C<NULL> or a pointer to an NV.
53305cf1 112The scan stops at the end of the string, or the first invalid character.
2d7f6611 113Unless C<PERL_SCAN_SILENT_ILLDIGIT> is set in C<*flags>, encountering an
7b667b5f 114invalid character will also trigger a warning.
2d7f6611
KW
115On return C<*len> is set to the length of the scanned string,
116and C<*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,
796b6530
KW
119and nothing is written to C<*result>. If the value is > C<UV_MAX>, C<grok_bin>
120returns C<UV_MAX>, sets C<PERL_SCAN_GREATER_THAN_UV_MAX> in the output flags,
2d7f6611 121and writes the value to C<*result> (or the value is discarded if C<result>
53305cf1
NC
122is NULL).
123
796b6530 124The binary number may optionally be prefixed with C<"0b"> or C<"b"> unless
2d7f6611
KW
125C<PERL_SCAN_DISALLOW_PREFIX> is set in C<*flags> on entry. If
126C<PERL_SCAN_ALLOW_UNDERSCORES> is set in C<*flags> then the binary
796b6530 127number may use C<"_"> characters to separate digits.
53305cf1
NC
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
2d7f6611 233On entry C<start> and C<*len_p> give the string to scan, C<*flags> gives
796b6530 234conversion flags, and C<result> should be C<NULL> or a pointer to an NV.
7b667b5f 235The scan stops at the end of the string, or the first invalid character.
2d7f6611 236Unless C<PERL_SCAN_SILENT_ILLDIGIT> is set in C<*flags>, encountering an
7b667b5f 237invalid character will also trigger a warning.
2d7f6611
KW
238On return C<*len> is set to the length of the scanned string,
239and C<*flags> gives output flags.
53305cf1 240
796b6530
KW
241If the value is <= C<UV_MAX> it is returned as a UV, the output flags are clear,
242and nothing is written to C<*result>. If the value is > C<UV_MAX>, C<grok_hex>
243returns C<UV_MAX>, sets C<PERL_SCAN_GREATER_THAN_UV_MAX> in the output flags,
2d7f6611 244and writes the value to C<*result> (or the value is discarded if C<result>
796b6530 245is C<NULL>).
53305cf1 246
796b6530 247The hex number may optionally be prefixed with C<"0x"> or C<"x"> unless
2d7f6611
KW
248C<PERL_SCAN_DISALLOW_PREFIX> is set in C<*flags> on entry. If
249C<PERL_SCAN_ALLOW_UNDERSCORES> is set in C<*flags> then the hex
796b6530 250number may use C<"_"> characters to separate digits.
53305cf1
NC
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
2d7f6611 354On entry C<start> and C<*len> give the string to scan, C<*flags> gives
796b6530 355conversion flags, and C<result> should be C<NULL> or a pointer to an NV.
7b667b5f 356The scan stops at the end of the string, or the first invalid character.
2d7f6611 357Unless C<PERL_SCAN_SILENT_ILLDIGIT> is set in C<*flags>, encountering an
154bd527 3588 or 9 will also trigger a warning.
2d7f6611
KW
359On return C<*len> is set to the length of the scanned string,
360and C<*flags> gives output flags.
7b667b5f 361
796b6530
KW
362If the value is <= C<UV_MAX> it is returned as a UV, the output flags are clear,
363and nothing is written to C<*result>. If the value is > C<UV_MAX>, C<grok_oct>
364returns C<UV_MAX>, sets C<PERL_SCAN_GREATER_THAN_UV_MAX> in the output flags,
2d7f6611 365and writes the value to C<*result> (or the value is discarded if C<result>
796b6530 366is C<NULL>).
7b667b5f 367
2d7f6611 368If C<PERL_SCAN_ALLOW_UNDERSCORES> is set in C<*flags> then the octal
796b6530 369number may use C<"_"> 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{
7918f24d
NC
521 PERL_ARGS_ASSERT_GROK_NUMERIC_RADIX;
522
7ea85fa8
KW
523#ifdef USE_LOCALE_NUMERIC
524
d6ded950 525 if (IN_LC(LC_NUMERIC)) {
67d796ae 526 DECLARATION_FOR_LC_NUMERIC_MANIPULATION;
a1395eaf 527 STORE_LC_NUMERIC_FORCE_TO_UNDERLYING();
3ca88433 528 {
c5a7e38e
KW
529 STRLEN len;
530 const char * const radix = SvPV(PL_numeric_radix_sv, len);
531 if (*sp + len <= send && memEQ(*sp, radix, len)) {
532 *sp += len;
533 RESTORE_LC_NUMERIC();
534 return TRUE;
535 }
21431899
KW
536 }
537 RESTORE_LC_NUMERIC();
98994639
HS
538 }
539 /* always try "." if numeric radix didn't match because
540 * we may have data from different locales mixed */
541#endif
7918f24d 542
98994639
HS
543 if (*sp < send && **sp == '.') {
544 ++*sp;
545 return TRUE;
546 }
547 return FALSE;
548}
549
569f27e5 550/*
ff4eb398
JH
551=for apidoc grok_infnan
552
796b6530 553Helper for C<grok_number()>, accepts various ways of spelling "infinity"
ff4eb398
JH
554or "not a number", and returns one of the following flag combinations:
555
556 IS_NUMBER_INFINITE
557 IS_NUMBER_NAN
558 IS_NUMBER_INFINITE | IS_NUMBER_NEG
559 IS_NUMBER_NAN | IS_NUMBER_NEG
560 0
561
796b6530 562possibly |-ed with C<IS_NUMBER_TRAILING>.
b489e20f 563
796b6530 564If an infinity or a not-a-number is recognized, C<*sp> will point to
62bdc035 565one byte past the end of the recognized string. If the recognition fails,
796b6530 566zero is returned, and C<*sp> will not move.
ff4eb398
JH
567
568=cut
569*/
570
571int
3823048b 572Perl_grok_infnan(pTHX_ const char** sp, const char* send)
ff4eb398
JH
573{
574 const char* s = *sp;
575 int flags = 0;
a5dc2484 576#if defined(NV_INF) || defined(NV_NAN)
62bdc035 577 bool odh = FALSE; /* one-dot-hash: 1.#INF */
ff4eb398
JH
578
579 PERL_ARGS_ASSERT_GROK_INFNAN;
580
8c12dc63
JH
581 if (*s == '+') {
582 s++; if (s == send) return 0;
583 }
584 else if (*s == '-') {
ff4eb398
JH
585 flags |= IS_NUMBER_NEG; /* Yes, -NaN happens. Incorrect but happens. */
586 s++; if (s == send) return 0;
587 }
588
589 if (*s == '1') {
62bdc035
JH
590 /* Visual C: 1.#SNAN, -1.#QNAN, 1#INF, 1.#IND (maybe also 1.#NAN)
591 * Let's keep the dot optional. */
ff4eb398
JH
592 s++; if (s == send) return 0;
593 if (*s == '.') {
594 s++; if (s == send) return 0;
595 }
596 if (*s == '#') {
597 s++; if (s == send) return 0;
598 } else
599 return 0;
e855f543 600 odh = TRUE;
ff4eb398
JH
601 }
602
305b8651 603 if (isALPHA_FOLD_EQ(*s, 'I')) {
62bdc035
JH
604 /* INF or IND (1.#IND is "indeterminate", a certain type of NAN) */
605
305b8651 606 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
ff4eb398 607 s++; if (s == send) return 0;
305b8651 608 if (isALPHA_FOLD_EQ(*s, 'F')) {
ff4eb398 609 s++;
b8974fcb
JH
610 if (s < send && (isALPHA_FOLD_EQ(*s, 'I'))) {
611 int fail =
612 flags | IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT | IS_NUMBER_TRAILING;
613 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return fail;
614 s++; if (s == send || isALPHA_FOLD_NE(*s, 'I')) return fail;
615 s++; if (s == send || isALPHA_FOLD_NE(*s, 'T')) return fail;
616 s++; if (s == send || isALPHA_FOLD_NE(*s, 'Y')) return fail;
3396ed30 617 s++;
b8974fcb
JH
618 } else if (odh) {
619 while (*s == '0') { /* 1.#INF00 */
620 s++;
621 }
3396ed30 622 }
b489e20f
JH
623 while (s < send && isSPACE(*s))
624 s++;
625 if (s < send && *s) {
3396ed30 626 flags |= IS_NUMBER_TRAILING;
fae4db12 627 }
ff4eb398
JH
628 flags |= IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT;
629 }
e855f543 630 else if (isALPHA_FOLD_EQ(*s, 'D') && odh) { /* 1.#IND */
ff4eb398
JH
631 s++;
632 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
fae4db12
JH
633 while (*s == '0') { /* 1.#IND00 */
634 s++;
635 }
1e9aa12f
JH
636 if (*s) {
637 flags |= IS_NUMBER_TRAILING;
638 }
ff4eb398
JH
639 } else
640 return 0;
ff4eb398
JH
641 }
642 else {
62bdc035 643 /* Maybe NAN of some sort */
3823048b
JH
644
645 if (isALPHA_FOLD_EQ(*s, 'S') || isALPHA_FOLD_EQ(*s, 'Q')) {
646 /* snan, qNaN */
647 /* XXX do something with the snan/qnan difference */
648 s++; if (s == send) return 0;
649 }
650
651 if (isALPHA_FOLD_EQ(*s, 'N')) {
652 s++; if (s == send || isALPHA_FOLD_NE(*s, 'A')) return 0;
653 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
654 s++;
655
656 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
657
658 /* NaN can be followed by various stuff (NaNQ, NaNS), but
659 * there are also multiple different NaN values, and some
660 * implementations output the "payload" values,
661 * e.g. NaN123, NAN(abc), while some legacy implementations
662 * have weird stuff like NaN%. */
663 if (isALPHA_FOLD_EQ(*s, 'q') ||
664 isALPHA_FOLD_EQ(*s, 's')) {
665 /* "nanq" or "nans" are ok, though generating
666 * these portably is tricky. */
667 s++;
668 }
669 if (*s == '(') {
670 /* C99 style "nan(123)" or Perlish equivalent "nan($uv)". */
671 const char *t;
672 s++;
673 if (s == send) {
674 return flags | IS_NUMBER_TRAILING;
675 }
676 t = s + 1;
677 while (t < send && *t && *t != ')') {
678 t++;
679 }
680 if (t == send) {
681 return flags | IS_NUMBER_TRAILING;
682 }
683 if (*t == ')') {
684 int nantype;
685 UV nanval;
686 if (s[0] == '0' && s + 2 < t &&
687 isALPHA_FOLD_EQ(s[1], 'x') &&
688 isXDIGIT(s[2])) {
689 STRLEN len = t - s;
690 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES;
691 nanval = grok_hex(s, &len, &flags, NULL);
692 if ((flags & PERL_SCAN_GREATER_THAN_UV_MAX)) {
693 nantype = 0;
694 } else {
695 nantype = IS_NUMBER_IN_UV;
696 }
697 s += len;
698 } else if (s[0] == '0' && s + 2 < t &&
699 isALPHA_FOLD_EQ(s[1], 'b') &&
700 (s[2] == '0' || s[2] == '1')) {
701 STRLEN len = t - s;
702 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES;
703 nanval = grok_bin(s, &len, &flags, NULL);
704 if ((flags & PERL_SCAN_GREATER_THAN_UV_MAX)) {
705 nantype = 0;
706 } else {
707 nantype = IS_NUMBER_IN_UV;
708 }
709 s += len;
710 } else {
711 const char *u;
712 nantype =
713 grok_number_flags(s, t - s, &nanval,
714 PERL_SCAN_TRAILING |
715 PERL_SCAN_ALLOW_UNDERSCORES);
716 /* Unfortunately grok_number_flags() doesn't
717 * tell how far we got and the ')' will always
718 * be "trailing", so we need to double-check
719 * whether we had something dubious. */
720 for (u = s; u < t; u++) {
721 if (!isDIGIT(*u)) {
722 flags |= IS_NUMBER_TRAILING;
723 break;
724 }
725 }
726 s = u;
727 }
728
729 /* XXX Doesn't do octal: nan("0123").
730 * Probably not a big loss. */
731
732 if ((nantype & IS_NUMBER_NOT_INT) ||
733 !(nantype && IS_NUMBER_IN_UV)) {
734 /* XXX the nanval is currently unused, that is,
735 * not inserted as the NaN payload of the NV.
736 * But the above code already parses the C99
737 * nan(...) format. See below, and see also
738 * the nan() in POSIX.xs.
739 *
740 * Certain configuration combinations where
741 * NVSIZE is greater than UVSIZE mean that
742 * a single UV cannot contain all the possible
743 * NaN payload bits. There would need to be
744 * some more generic syntax than "nan($uv)".
745 *
746 * Issues to keep in mind:
747 *
748 * (1) In most common cases there would
749 * not be an integral number of bytes that
750 * could be set, only a certain number of bits.
751 * For example for the common case of
752 * NVSIZE == UVSIZE == 8 there is room for 52
753 * bits in the payload, but the most significant
754 * bit is commonly reserved for the
755 * signaling/quiet bit, leaving 51 bits.
756 * Furthermore, the C99 nan() is supposed
757 * to generate quiet NaNs, so it is doubtful
758 * whether it should be able to generate
759 * signaling NaNs. For the x86 80-bit doubles
760 * (if building a long double Perl) there would
761 * be 62 bits (s/q bit being the 63rd).
762 *
763 * (2) Endianness of the payload bits. If the
764 * payload is specified as an UV, the low-order
765 * bits of the UV are naturally little-endianed
766 * (rightmost) bits of the payload. The endianness
767 * of UVs and NVs can be different. */
768 return 0;
769 }
770 if (s < t) {
771 flags |= IS_NUMBER_TRAILING;
772 }
773 } else {
774 /* Looked like nan(...), but no close paren. */
775 flags |= IS_NUMBER_TRAILING;
776 }
777 } else {
778 while (s < send && isSPACE(*s))
779 s++;
780 if (s < send && *s) {
781 /* Note that we here implicitly accept (parse as
782 * "nan", but with warnings) also any other weird
783 * trailing stuff for "nan". In the above we just
784 * check that if we got the C99-style "nan(...)",
785 * the "..." looks sane.
786 * If in future we accept more ways of specifying
787 * the nan payload, the accepting would happen around
788 * here. */
789 flags |= IS_NUMBER_TRAILING;
790 }
791 }
792 s = send;
793 }
794 else
795 return 0;
ff4eb398
JH
796 }
797
b489e20f
JH
798 while (s < send && isSPACE(*s))
799 s++;
800
a5dc2484
JH
801#else
802 PERL_UNUSED_ARG(send);
803#endif /* #if defined(NV_INF) || defined(NV_NAN) */
a1fe7cea
JH
804 *sp = s;
805 return flags;
ff4eb398
JH
806}
807
13393a5e 808/*
3823048b 809=for apidoc grok_number_flags
13393a5e
JH
810
811Recognise (or not) a number. The type of the number is returned
812(0 if unrecognised), otherwise it is a bit-ORed combination of
796b6530
KW
813C<IS_NUMBER_IN_UV>, C<IS_NUMBER_GREATER_THAN_UV_MAX>, C<IS_NUMBER_NOT_INT>,
814C<IS_NUMBER_NEG>, C<IS_NUMBER_INFINITY>, C<IS_NUMBER_NAN> (defined in perl.h).
815
816If the value of the number can fit in a UV, it is returned in C<*valuep>.
817C<IS_NUMBER_IN_UV> will be set to indicate that C<*valuep> is valid, C<IS_NUMBER_IN_UV>
818will never be set unless C<*valuep> is valid, but C<*valuep> may have been assigned
819to during processing even though C<IS_NUMBER_IN_UV> is not set on return.
820If C<valuep> is C<NULL>, C<IS_NUMBER_IN_UV> will be set for the same cases as when
821C<valuep> is non-C<NULL>, but no actual assignment (or SEGV) will occur.
822
823C<IS_NUMBER_NOT_INT> will be set with C<IS_NUMBER_IN_UV> if trailing decimals were
824seen (in which case C<*valuep> gives the true value truncated to an integer), and
825C<IS_NUMBER_NEG> if the number is negative (in which case C<*valuep> holds the
826absolute value). C<IS_NUMBER_IN_UV> is not set if e notation was used or the
13393a5e
JH
827number is larger than a UV.
828
829C<flags> allows only C<PERL_SCAN_TRAILING>, which allows for trailing
830non-numeric text on an otherwise successful I<grok>, setting
831C<IS_NUMBER_TRAILING> on the result.
832
833=for apidoc grok_number
834
796b6530 835Identical to C<grok_number_flags()> with C<flags> set to zero.
13393a5e
JH
836
837=cut
838 */
839int
840Perl_grok_number(pTHX_ const char *pv, STRLEN len, UV *valuep)
841{
842 PERL_ARGS_ASSERT_GROK_NUMBER;
843
844 return grok_number_flags(pv, len, valuep, 0);
845}
846
945b524a
JH
847static const UV uv_max_div_10 = UV_MAX / 10;
848static const U8 uv_max_mod_10 = UV_MAX % 10;
849
3f7602fa 850int
3823048b 851Perl_grok_number_flags(pTHX_ const char *pv, STRLEN len, UV *valuep, U32 flags)
3f7602fa 852{
60939fb8 853 const char *s = pv;
c4420975 854 const char * const send = pv + len;
ae776a2c 855 const char *d;
60939fb8 856 int numtype = 0;
60939fb8 857
3823048b 858 PERL_ARGS_ASSERT_GROK_NUMBER_FLAGS;
7918f24d 859
60939fb8
NC
860 while (s < send && isSPACE(*s))
861 s++;
862 if (s == send) {
863 return 0;
864 } else if (*s == '-') {
865 s++;
866 numtype = IS_NUMBER_NEG;
867 }
868 else if (*s == '+')
aa42a541 869 s++;
60939fb8
NC
870
871 if (s == send)
872 return 0;
873
ae776a2c 874 /* The first digit (after optional sign): note that might
8c12dc63 875 * also point to "infinity" or "nan", or "1.#INF". */
ae776a2c
JH
876 d = s;
877
8c12dc63 878 /* next must be digit or the radix separator or beginning of infinity/nan */
60939fb8
NC
879 if (isDIGIT(*s)) {
880 /* UVs are at least 32 bits, so the first 9 decimal digits cannot
881 overflow. */
882 UV value = *s - '0';
883 /* This construction seems to be more optimiser friendly.
884 (without it gcc does the isDIGIT test and the *s - '0' separately)
885 With it gcc on arm is managing 6 instructions (6 cycles) per digit.
886 In theory the optimiser could deduce how far to unroll the loop
887 before checking for overflow. */
58bb9ec3
NC
888 if (++s < send) {
889 int digit = *s - '0';
60939fb8
NC
890 if (digit >= 0 && digit <= 9) {
891 value = value * 10 + digit;
58bb9ec3
NC
892 if (++s < send) {
893 digit = *s - '0';
60939fb8
NC
894 if (digit >= 0 && digit <= 9) {
895 value = value * 10 + digit;
58bb9ec3
NC
896 if (++s < send) {
897 digit = *s - '0';
60939fb8
NC
898 if (digit >= 0 && digit <= 9) {
899 value = value * 10 + digit;
58bb9ec3
NC
900 if (++s < send) {
901 digit = *s - '0';
60939fb8
NC
902 if (digit >= 0 && digit <= 9) {
903 value = value * 10 + digit;
58bb9ec3
NC
904 if (++s < send) {
905 digit = *s - '0';
60939fb8
NC
906 if (digit >= 0 && digit <= 9) {
907 value = value * 10 + digit;
58bb9ec3
NC
908 if (++s < send) {
909 digit = *s - '0';
60939fb8
NC
910 if (digit >= 0 && digit <= 9) {
911 value = value * 10 + digit;
58bb9ec3
NC
912 if (++s < send) {
913 digit = *s - '0';
60939fb8
NC
914 if (digit >= 0 && digit <= 9) {
915 value = value * 10 + digit;
58bb9ec3
NC
916 if (++s < send) {
917 digit = *s - '0';
60939fb8
NC
918 if (digit >= 0 && digit <= 9) {
919 value = value * 10 + digit;
58bb9ec3 920 if (++s < send) {
60939fb8
NC
921 /* Now got 9 digits, so need to check
922 each time for overflow. */
58bb9ec3 923 digit = *s - '0';
60939fb8 924 while (digit >= 0 && digit <= 9
945b524a
JH
925 && (value < uv_max_div_10
926 || (value == uv_max_div_10
927 && digit <= uv_max_mod_10))) {
60939fb8 928 value = value * 10 + digit;
58bb9ec3
NC
929 if (++s < send)
930 digit = *s - '0';
60939fb8
NC
931 else
932 break;
933 }
934 if (digit >= 0 && digit <= 9
51bd16da 935 && (s < send)) {
60939fb8
NC
936 /* value overflowed.
937 skip the remaining digits, don't
938 worry about setting *valuep. */
939 do {
940 s++;
941 } while (s < send && isDIGIT(*s));
942 numtype |=
943 IS_NUMBER_GREATER_THAN_UV_MAX;
944 goto skip_value;
945 }
946 }
947 }
98994639 948 }
60939fb8
NC
949 }
950 }
951 }
952 }
953 }
954 }
955 }
956 }
957 }
958 }
959 }
98994639 960 }
60939fb8 961 }
98994639 962 }
60939fb8
NC
963 numtype |= IS_NUMBER_IN_UV;
964 if (valuep)
965 *valuep = value;
966
967 skip_value:
968 if (GROK_NUMERIC_RADIX(&s, send)) {
969 numtype |= IS_NUMBER_NOT_INT;
970 while (s < send && isDIGIT(*s)) /* optional digits after the radix */
971 s++;
98994639 972 }
60939fb8
NC
973 }
974 else if (GROK_NUMERIC_RADIX(&s, send)) {
975 numtype |= IS_NUMBER_NOT_INT | IS_NUMBER_IN_UV; /* valuep assigned below */
976 /* no digits before the radix means we need digits after it */
977 if (s < send && isDIGIT(*s)) {
978 do {
979 s++;
980 } while (s < send && isDIGIT(*s));
981 if (valuep) {
982 /* integer approximation is valid - it's 0. */
983 *valuep = 0;
984 }
98994639 985 }
60939fb8 986 else
ae776a2c 987 return 0;
ff4eb398 988 }
60939fb8 989
926f5fc6 990 if (s > d && s < send) {
60939fb8 991 /* we can have an optional exponent part */
305b8651 992 if (isALPHA_FOLD_EQ(*s, 'e')) {
60939fb8
NC
993 s++;
994 if (s < send && (*s == '-' || *s == '+'))
995 s++;
996 if (s < send && isDIGIT(*s)) {
997 do {
998 s++;
999 } while (s < send && isDIGIT(*s));
1000 }
3f7602fa
TC
1001 else if (flags & PERL_SCAN_TRAILING)
1002 return numtype | IS_NUMBER_TRAILING;
60939fb8 1003 else
3f7602fa
TC
1004 return 0;
1005
1006 /* The only flag we keep is sign. Blow away any "it's UV" */
1007 numtype &= IS_NUMBER_NEG;
1008 numtype |= IS_NUMBER_NOT_INT;
60939fb8
NC
1009 }
1010 }
1011 while (s < send && isSPACE(*s))
1012 s++;
1013 if (s >= send)
aa8b85de 1014 return numtype;
b59bf0b2 1015 if (memEQs(pv, len, "0 but true")) {
60939fb8
NC
1016 if (valuep)
1017 *valuep = 0;
1018 return IS_NUMBER_IN_UV;
1019 }
8c12dc63
JH
1020 /* We could be e.g. at "Inf" or "NaN", or at the "#" of "1.#INF". */
1021 if ((s + 2 < send) && strchr("inqs#", toFOLD(*s))) {
1022 /* Really detect inf/nan. Start at d, not s, since the above
1023 * code might have already consumed the "1." or "1". */
7eff3d39 1024 const int infnan = Perl_grok_infnan(aTHX_ &d, send);
8c12dc63
JH
1025 if ((infnan & IS_NUMBER_INFINITY)) {
1026 return (numtype | infnan); /* Keep sign for infinity. */
1027 }
1028 else if ((infnan & IS_NUMBER_NAN)) {
1029 return (numtype | infnan) & ~IS_NUMBER_NEG; /* Clear sign for nan. */
1030 }
1031 }
3f7602fa
TC
1032 else if (flags & PERL_SCAN_TRAILING) {
1033 return numtype | IS_NUMBER_TRAILING;
1034 }
1035
60939fb8 1036 return 0;
98994639
HS
1037}
1038
6313e544 1039/*
73e43954 1040grok_atoUV
6313e544 1041
22ff3130 1042grok_atoUV parses a C-style zero-byte terminated string, looking for
d62b8c6a 1043a decimal unsigned integer.
338aa8b0 1044
d62b8c6a
JH
1045Returns the unsigned integer, if a valid value can be parsed
1046from the beginning of the string.
f4379102 1047
d62b8c6a 1048Accepts only the decimal digits '0'..'9'.
6313e544 1049
22ff3130 1050As opposed to atoi or strtol, grok_atoUV does NOT allow optional
d62b8c6a
JH
1051leading whitespace, or negative inputs. If such features are
1052required, the calling code needs to explicitly implement those.
6313e544 1053
22ff3130
HS
1054Returns true if a valid value could be parsed. In that case, valptr
1055is set to the parsed value, and endptr (if provided) is set to point
1056to the character after the last digit.
338aa8b0 1057
22ff3130
HS
1058Returns false otherwise. This can happen if a) there is a leading zero
1059followed by another digit; b) the digits would overflow a UV; or c)
1060there are trailing non-digits AND endptr is not provided.
6313e544 1061
d62b8c6a
JH
1062Background: atoi has severe problems with illegal inputs, it cannot be
1063used for incremental parsing, and therefore should be avoided
1064atoi and strtol are also affected by locale settings, which can also be
1065seen as a bug (global state controlled by user environment).
1066
6313e544
JH
1067*/
1068
22ff3130
HS
1069bool
1070Perl_grok_atoUV(const char *pv, UV *valptr, const char** endptr)
6313e544
JH
1071{
1072 const char* s = pv;
1073 const char** eptr;
1074 const char* end2; /* Used in case endptr is NULL. */
22ff3130 1075 UV val = 0; /* The parsed value. */
6313e544 1076
22ff3130 1077 PERL_ARGS_ASSERT_GROK_ATOUV;
6313e544
JH
1078
1079 eptr = endptr ? endptr : &end2;
75feedba
JH
1080 if (isDIGIT(*s)) {
1081 /* Single-digit inputs are quite common. */
6313e544 1082 val = *s++ - '0';
75feedba 1083 if (isDIGIT(*s)) {
22ff3130
HS
1084 /* Fail on extra leading zeros. */
1085 if (val == 0)
1086 return FALSE;
75feedba
JH
1087 while (isDIGIT(*s)) {
1088 /* This could be unrolled like in grok_number(), but
1089 * the expected uses of this are not speed-needy, and
1090 * unlikely to need full 64-bitness. */
7eff3d39 1091 const U8 digit = *s++ - '0';
945b524a
JH
1092 if (val < uv_max_div_10 ||
1093 (val == uv_max_div_10 && digit <= uv_max_mod_10)) {
75feedba
JH
1094 val = val * 10 + digit;
1095 } else {
22ff3130 1096 return FALSE;
6313e544 1097 }
6313e544
JH
1098 }
1099 }
75feedba 1100 }
22ff3130
HS
1101 if (s == pv)
1102 return FALSE;
1103 if (endptr == NULL && *s)
1104 return FALSE; /* If endptr is NULL, no trailing non-digits allowed. */
6313e544 1105 *eptr = s;
22ff3130
HS
1106 *valptr = val;
1107 return TRUE;
6313e544
JH
1108}
1109
a4eca1d4 1110#ifndef USE_QUADMATH
4801ca72 1111STATIC NV
98994639
HS
1112S_mulexp10(NV value, I32 exponent)
1113{
1114 NV result = 1.0;
1115 NV power = 10.0;
1116 bool negative = 0;
1117 I32 bit;
1118
1119 if (exponent == 0)
1120 return value;
659c4b96
DM
1121 if (value == 0)
1122 return (NV)0;
87032ba1 1123
24866caa 1124 /* On OpenVMS VAX we by default use the D_FLOAT double format,
67597c89 1125 * and that format does not have *easy* capabilities [1] for
24866caa
CB
1126 * overflowing doubles 'silently' as IEEE fp does. We also need
1127 * to support G_FLOAT on both VAX and Alpha, and though the exponent
1128 * range is much larger than D_FLOAT it still doesn't do silent
1129 * overflow. Therefore we need to detect early whether we would
1130 * overflow (this is the behaviour of the native string-to-float
1131 * conversion routines, and therefore of native applications, too).
67597c89 1132 *
24866caa
CB
1133 * [1] Trying to establish a condition handler to trap floating point
1134 * exceptions is not a good idea. */
87032ba1
JH
1135
1136 /* In UNICOS and in certain Cray models (such as T90) there is no
1137 * IEEE fp, and no way at all from C to catch fp overflows gracefully.
1138 * There is something you can do if you are willing to use some
1139 * inline assembler: the instruction is called DFI-- but that will
1140 * disable *all* floating point interrupts, a little bit too large
1141 * a hammer. Therefore we need to catch potential overflows before
1142 * it's too late. */
353813d9 1143
a7157111 1144#if ((defined(VMS) && !defined(_IEEE_FP)) || defined(_UNICOS) || defined(DOUBLE_IS_VAX_FLOAT)) && defined(NV_MAX_10_EXP)
353813d9 1145 STMT_START {
c4420975 1146 const NV exp_v = log10(value);
353813d9
HS
1147 if (exponent >= NV_MAX_10_EXP || exponent + exp_v >= NV_MAX_10_EXP)
1148 return NV_MAX;
1149 if (exponent < 0) {
1150 if (-(exponent + exp_v) >= NV_MAX_10_EXP)
1151 return 0.0;
1152 while (-exponent >= NV_MAX_10_EXP) {
1153 /* combination does not overflow, but 10^(-exponent) does */
1154 value /= 10;
1155 ++exponent;
1156 }
1157 }
1158 } STMT_END;
87032ba1
JH
1159#endif
1160
353813d9
HS
1161 if (exponent < 0) {
1162 negative = 1;
1163 exponent = -exponent;
b27804d8
DM
1164#ifdef NV_MAX_10_EXP
1165 /* for something like 1234 x 10^-309, the action of calculating
1166 * the intermediate value 10^309 then returning 1234 / (10^309)
1167 * will fail, since 10^309 becomes infinity. In this case try to
1168 * refactor it as 123 / (10^308) etc.
1169 */
1170 while (value && exponent > NV_MAX_10_EXP) {
1171 exponent--;
1172 value /= 10;
1173 }
48853916
JH
1174 if (value == 0.0)
1175 return value;
b27804d8 1176#endif
353813d9 1177 }
c62e754c
JH
1178#if defined(__osf__)
1179 /* Even with cc -ieee + ieee_set_fp_control(IEEE_TRAP_ENABLE_INV)
1180 * Tru64 fp behavior on inf/nan is somewhat broken. Another way
1181 * to do this would be ieee_set_fp_control(IEEE_TRAP_ENABLE_OVF)
1182 * but that breaks another set of infnan.t tests. */
1183# define FP_OVERFLOWS_TO_ZERO
1184#endif
98994639
HS
1185 for (bit = 1; exponent; bit <<= 1) {
1186 if (exponent & bit) {
1187 exponent ^= bit;
1188 result *= power;
c62e754c
JH
1189#ifdef FP_OVERFLOWS_TO_ZERO
1190 if (result == 0)
a7157111 1191# ifdef NV_INF
c62e754c 1192 return value < 0 ? -NV_INF : NV_INF;
a7157111
JH
1193# else
1194 return value < 0 ? -FLT_MAX : FLT_MAX;
1195# endif
c62e754c 1196#endif
236f0012
CB
1197 /* Floating point exceptions are supposed to be turned off,
1198 * but if we're obviously done, don't risk another iteration.
1199 */
1200 if (exponent == 0) break;
98994639
HS
1201 }
1202 power *= power;
1203 }
1204 return negative ? value / result : value * result;
1205}
a4eca1d4 1206#endif /* #ifndef USE_QUADMATH */
98994639
HS
1207
1208NV
1209Perl_my_atof(pTHX_ const char* s)
1210{
f720c878
KW
1211 /* 's' must be NUL terminated */
1212
98994639 1213 NV x = 0.0;
9eda1ea6
KW
1214
1215 PERL_ARGS_ASSERT_MY_ATOF;
1216
a4eca1d4 1217#ifdef USE_QUADMATH
9eda1ea6 1218
a4eca1d4 1219 Perl_my_atof2(aTHX_ s, &x);
9eda1ea6
KW
1220
1221#elif ! defined(USE_LOCALE_NUMERIC)
1222
1223 Perl_atof2(s, x);
1224
1225#else
7918f24d 1226
a2287a13 1227 {
67d796ae
KW
1228 DECLARATION_FOR_LC_NUMERIC_MANIPULATION;
1229 STORE_LC_NUMERIC_SET_TO_NEEDED();
d6ded950 1230 if (PL_numeric_radix_sv && IN_LC(LC_NUMERIC)) {
e4850248
KW
1231 /* Look through the string for the first thing that looks like a
1232 * decimal point: either the value in the current locale or the
1233 * standard fallback of '.'. The one which appears earliest in the
1234 * input string is the one that we should have atof look for. Note
1235 * that we have to determine this beforehand because on some
1236 * systems, Perl_atof2 is just a wrapper around the system's atof.
1237 * */
1ae85f6c
KW
1238 const char * const standard_pos = strchr(s, '.');
1239 const char * const local_pos
1240 = strstr(s, SvPV_nolen(PL_numeric_radix_sv));
1241 const bool use_standard_radix
1242 = standard_pos && (!local_pos || standard_pos < local_pos);
78787052 1243
665873e9 1244 if (use_standard_radix) {
e4850248 1245 SET_NUMERIC_STANDARD();
665873e9
KW
1246 LOCK_LC_NUMERIC_STANDARD();
1247 }
78787052 1248
e4850248 1249 Perl_atof2(s, x);
78787052 1250
665873e9
KW
1251 if (use_standard_radix) {
1252 UNLOCK_LC_NUMERIC_STANDARD();
67d796ae 1253 SET_NUMERIC_UNDERLYING();
665873e9 1254 }
e4850248
KW
1255 }
1256 else
1257 Perl_atof2(s, x);
a2287a13
KW
1258 RESTORE_LC_NUMERIC();
1259 }
9eda1ea6 1260
98994639 1261#endif
9eda1ea6 1262
98994639
HS
1263 return x;
1264}
1265
a7157111 1266#if defined(NV_INF) || defined(NV_NAN)
3c81f0b3
DD
1267
1268#ifdef USING_MSVC6
1269# pragma warning(push)
1270# pragma warning(disable:4756;disable:4056)
1271#endif
829757a4 1272static char*
5563f457 1273S_my_atof_infnan(pTHX_ const char* s, bool negative, const char* send, NV* value)
829757a4
JH
1274{
1275 const char *p0 = negative ? s - 1 : s;
1276 const char *p = p0;
7eff3d39 1277 const int infnan = grok_infnan(&p, send);
829757a4
JH
1278 if (infnan && p != p0) {
1279 /* If we can generate inf/nan directly, let's do so. */
1280#ifdef NV_INF
1281 if ((infnan & IS_NUMBER_INFINITY)) {
3823048b 1282 *value = (infnan & IS_NUMBER_NEG) ? -NV_INF: NV_INF;
829757a4
JH
1283 return (char*)p;
1284 }
1285#endif
1286#ifdef NV_NAN
1287 if ((infnan & IS_NUMBER_NAN)) {
3823048b 1288 *value = NV_NAN;
829757a4
JH
1289 return (char*)p;
1290 }
1291#endif
1292#ifdef Perl_strtod
68611e6f 1293 /* If still here, we didn't have either NV_INF or NV_NAN,
829757a4
JH
1294 * and can try falling back to native strtod/strtold.
1295 *
1296 * The native interface might not recognize all the possible
1297 * inf/nan strings Perl recognizes. What we can try
1298 * is to try faking the input. We will try inf/-inf/nan
1299 * as the most promising/portable input. */
1300 {
1301 const char* fake = NULL;
1302 char* endp;
1303 NV nv;
a7157111 1304#ifdef NV_INF
829757a4
JH
1305 if ((infnan & IS_NUMBER_INFINITY)) {
1306 fake = ((infnan & IS_NUMBER_NEG)) ? "-inf" : "inf";
1307 }
a7157111
JH
1308#endif
1309#ifdef NV_NAN
1310 if ((infnan & IS_NUMBER_NAN)) {
829757a4
JH
1311 fake = "nan";
1312 }
a7157111 1313#endif
829757a4
JH
1314 assert(fake);
1315 nv = Perl_strtod(fake, &endp);
1316 if (fake != endp) {
a7157111 1317#ifdef NV_INF
829757a4 1318 if ((infnan & IS_NUMBER_INFINITY)) {
a7157111 1319# ifdef Perl_isinf
829757a4
JH
1320 if (Perl_isinf(nv))
1321 *value = nv;
a7157111 1322# else
829757a4
JH
1323 /* last resort, may generate SIGFPE */
1324 *value = Perl_exp((NV)1e9);
1325 if ((infnan & IS_NUMBER_NEG))
1326 *value = -*value;
a7157111 1327# endif
829757a4
JH
1328 return (char*)p; /* p, not endp */
1329 }
a7157111
JH
1330#endif
1331#ifdef NV_NAN
1332 if ((infnan & IS_NUMBER_NAN)) {
1333# ifdef Perl_isnan
829757a4
JH
1334 if (Perl_isnan(nv))
1335 *value = nv;
a7157111 1336# else
829757a4
JH
1337 /* last resort, may generate SIGFPE */
1338 *value = Perl_log((NV)-1.0);
a7157111 1339# endif
829757a4 1340 return (char*)p; /* p, not endp */
a7157111 1341#endif
829757a4
JH
1342 }
1343 }
1344 }
1345#endif /* #ifdef Perl_strtod */
1346 }
1347 return NULL;
1348}
3c81f0b3
DD
1349#ifdef USING_MSVC6
1350# pragma warning(pop)
1351#endif
829757a4 1352
a7157111
JH
1353#endif /* if defined(NV_INF) || defined(NV_NAN) */
1354
98994639
HS
1355char*
1356Perl_my_atof2(pTHX_ const char* orig, NV* value)
1357{
e1ec3a88 1358 const char* s = orig;
a4eca1d4
JH
1359 NV result[3] = {0.0, 0.0, 0.0};
1360#if defined(USE_PERL_ATOF) || defined(USE_QUADMATH)
ae776a2c 1361 const char* send = s + strlen(orig); /* one past the last */
a4eca1d4
JH
1362 bool negative = 0;
1363#endif
1364#if defined(USE_PERL_ATOF) && !defined(USE_QUADMATH)
1365 UV accumulator[2] = {0,0}; /* before/after dp */
8194bf88 1366 bool seen_digit = 0;
20f6aaab
AS
1367 I32 exp_adjust[2] = {0,0};
1368 I32 exp_acc[2] = {-1, -1};
1369 /* the current exponent adjust for the accumulators */
98994639 1370 I32 exponent = 0;
8194bf88 1371 I32 seen_dp = 0;
20f6aaab
AS
1372 I32 digit = 0;
1373 I32 old_digit = 0;
8194bf88 1374 I32 sig_digits = 0; /* noof significant digits seen so far */
a4eca1d4 1375#endif
8194bf88 1376
a4eca1d4 1377#if defined(USE_PERL_ATOF) || defined(USE_QUADMATH)
7918f24d
NC
1378 PERL_ARGS_ASSERT_MY_ATOF2;
1379
a4eca1d4
JH
1380 /* leading whitespace */
1381 while (isSPACE(*s))
1382 ++s;
1383
1384 /* sign */
1385 switch (*s) {
1386 case '-':
1387 negative = 1;
1388 /* FALLTHROUGH */
1389 case '+':
1390 ++s;
1391 }
1392#endif
1393
1394#ifdef USE_QUADMATH
1395 {
1396 char* endp;
adc55e02 1397 if ((endp = S_my_atof_infnan(aTHX_ s, negative, send, value)))
a4eca1d4
JH
1398 return endp;
1399 result[2] = strtoflt128(s, &endp);
1400 if (s != endp) {
1401 *value = negative ? -result[2] : result[2];
1402 return endp;
1403 }
1404 return NULL;
1405 }
1406#elif defined(USE_PERL_ATOF)
1407
8194bf88
DM
1408/* There is no point in processing more significant digits
1409 * than the NV can hold. Note that NV_DIG is a lower-bound value,
1410 * while we need an upper-bound value. We add 2 to account for this;
1411 * since it will have been conservative on both the first and last digit.
1412 * For example a 32-bit mantissa with an exponent of 4 would have
1413 * exact values in the set
1414 * 4
1415 * 8
1416 * ..
1417 * 17179869172
1418 * 17179869176
1419 * 17179869180
1420 *
1421 * where for the purposes of calculating NV_DIG we would have to discount
1422 * both the first and last digit, since neither can hold all values from
1423 * 0..9; but for calculating the value we must examine those two digits.
1424 */
ffa277e5
AS
1425#ifdef MAX_SIG_DIG_PLUS
1426 /* It is not necessarily the case that adding 2 to NV_DIG gets all the
1427 possible digits in a NV, especially if NVs are not IEEE compliant
1428 (e.g., long doubles on IRIX) - Allen <allens@cpan.org> */
1429# define MAX_SIG_DIGITS (NV_DIG+MAX_SIG_DIG_PLUS)
1430#else
1431# define MAX_SIG_DIGITS (NV_DIG+2)
1432#endif
8194bf88
DM
1433
1434/* the max number we can accumulate in a UV, and still safely do 10*N+9 */
1435#define MAX_ACCUMULATE ( (UV) ((UV_MAX - 9)/10))
98994639 1436
a5dc2484 1437#if defined(NV_INF) || defined(NV_NAN)
ae776a2c 1438 {
7eff3d39 1439 char* endp;
5563f457 1440 if ((endp = S_my_atof_infnan(aTHX_ s, negative, send, value)))
7eff3d39 1441 return endp;
ae776a2c 1442 }
a5dc2484 1443#endif
2b54f59f 1444
8194bf88
DM
1445 /* we accumulate digits into an integer; when this becomes too
1446 * large, we add the total to NV and start again */
98994639 1447
8194bf88
DM
1448 while (1) {
1449 if (isDIGIT(*s)) {
1450 seen_digit = 1;
20f6aaab 1451 old_digit = digit;
8194bf88 1452 digit = *s++ - '0';
20f6aaab
AS
1453 if (seen_dp)
1454 exp_adjust[1]++;
98994639 1455
8194bf88
DM
1456 /* don't start counting until we see the first significant
1457 * digit, eg the 5 in 0.00005... */
1458 if (!sig_digits && digit == 0)
1459 continue;
1460
1461 if (++sig_digits > MAX_SIG_DIGITS) {
98994639 1462 /* limits of precision reached */
20f6aaab
AS
1463 if (digit > 5) {
1464 ++accumulator[seen_dp];
1465 } else if (digit == 5) {
1466 if (old_digit % 2) { /* round to even - Allen */
1467 ++accumulator[seen_dp];
1468 }
1469 }
1470 if (seen_dp) {
1471 exp_adjust[1]--;
1472 } else {
1473 exp_adjust[0]++;
1474 }
8194bf88 1475 /* skip remaining digits */
98994639 1476 while (isDIGIT(*s)) {
98994639 1477 ++s;
20f6aaab
AS
1478 if (! seen_dp) {
1479 exp_adjust[0]++;
1480 }
98994639
HS
1481 }
1482 /* warn of loss of precision? */
98994639 1483 }
8194bf88 1484 else {
20f6aaab 1485 if (accumulator[seen_dp] > MAX_ACCUMULATE) {
8194bf88 1486 /* add accumulator to result and start again */
20f6aaab
AS
1487 result[seen_dp] = S_mulexp10(result[seen_dp],
1488 exp_acc[seen_dp])
1489 + (NV)accumulator[seen_dp];
1490 accumulator[seen_dp] = 0;
1491 exp_acc[seen_dp] = 0;
98994639 1492 }
20f6aaab
AS
1493 accumulator[seen_dp] = accumulator[seen_dp] * 10 + digit;
1494 ++exp_acc[seen_dp];
98994639 1495 }
8194bf88 1496 }
e1ec3a88 1497 else if (!seen_dp && GROK_NUMERIC_RADIX(&s, send)) {
8194bf88 1498 seen_dp = 1;
20f6aaab 1499 if (sig_digits > MAX_SIG_DIGITS) {
9604fbf0 1500 while (isDIGIT(*s)) {
20f6aaab 1501 ++s;
9604fbf0 1502 }
20f6aaab
AS
1503 break;
1504 }
8194bf88
DM
1505 }
1506 else {
1507 break;
98994639
HS
1508 }
1509 }
1510
20f6aaab
AS
1511 result[0] = S_mulexp10(result[0], exp_acc[0]) + (NV)accumulator[0];
1512 if (seen_dp) {
1513 result[1] = S_mulexp10(result[1], exp_acc[1]) + (NV)accumulator[1];
1514 }
98994639 1515
305b8651 1516 if (seen_digit && (isALPHA_FOLD_EQ(*s, 'e'))) {
98994639
HS
1517 bool expnegative = 0;
1518
1519 ++s;
1520 switch (*s) {
1521 case '-':
1522 expnegative = 1;
924ba076 1523 /* FALLTHROUGH */
98994639
HS
1524 case '+':
1525 ++s;
1526 }
1527 while (isDIGIT(*s))
1528 exponent = exponent * 10 + (*s++ - '0');
1529 if (expnegative)
1530 exponent = -exponent;
1531 }
1532
20f6aaab
AS
1533
1534
98994639 1535 /* now apply the exponent */
20f6aaab
AS
1536
1537 if (seen_dp) {
1538 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0])
1539 + S_mulexp10(result[1],exponent-exp_adjust[1]);
1540 } else {
1541 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0]);
1542 }
98994639
HS
1543
1544 /* now apply the sign */
1545 if (negative)
20f6aaab 1546 result[2] = -result[2];
a36244b7 1547#endif /* USE_PERL_ATOF */
20f6aaab 1548 *value = result[2];
73d840c0 1549 return (char *)s;
98994639
HS
1550}
1551
5d34af89 1552/*
3d9d9213 1553=for apidoc isinfnan
5d34af89 1554
796b6530
KW
1555C<Perl_isinfnan()> is utility function that returns true if the NV
1556argument is either an infinity or a C<NaN>, false otherwise. To test
1557in more detail, use C<Perl_isinf()> and C<Perl_isnan()>.
5d34af89 1558
68611e6f
JH
1559This is also the logical inverse of Perl_isfinite().
1560
5d34af89
JH
1561=cut
1562*/
1cd88304
JH
1563bool
1564Perl_isinfnan(NV nv)
1565{
a5dc2484 1566 PERL_UNUSED_ARG(nv);
1cd88304
JH
1567#ifdef Perl_isinf
1568 if (Perl_isinf(nv))
1569 return TRUE;
1570#endif
1571#ifdef Perl_isnan
1572 if (Perl_isnan(nv))
1573 return TRUE;
1574#endif
1575 return FALSE;
1576}
1577
354b74ae
FC
1578/*
1579=for apidoc
1580
796b6530 1581Checks whether the argument would be either an infinity or C<NaN> when used
354b74ae 1582as a number, but is careful not to trigger non-numeric or uninitialized
796b6530 1583warnings. it assumes the caller has done C<SvGETMAGIC(sv)> already.
354b74ae
FC
1584
1585=cut
1586*/
1587
1588bool
1589Perl_isinfnansv(pTHX_ SV *sv)
1590{
1591 PERL_ARGS_ASSERT_ISINFNANSV;
1592 if (!SvOK(sv))
1593 return FALSE;
1594 if (SvNOKp(sv))
1595 return Perl_isinfnan(SvNVX(sv));
1596 if (SvIOKp(sv))
1597 return FALSE;
1598 {
1599 STRLEN len;
1600 const char *s = SvPV_nomg_const(sv, len);
3823048b 1601 return cBOOL(grok_infnan(&s, s+len));
354b74ae
FC
1602 }
1603}
1604
d67dac15 1605#ifndef HAS_MODFL
68611e6f
JH
1606/* C99 has truncl, pre-C99 Solaris had aintl. We can use either with
1607 * copysignl to emulate modfl, which is in some platforms missing or
1608 * broken. */
d67dac15
JH
1609# if defined(HAS_TRUNCL) && defined(HAS_COPYSIGNL)
1610long double
1611Perl_my_modfl(long double x, long double *ip)
1612{
68611e6f
JH
1613 *ip = truncl(x);
1614 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
d67dac15
JH
1615}
1616# elif defined(HAS_AINTL) && defined(HAS_COPYSIGNL)
55954f19
JH
1617long double
1618Perl_my_modfl(long double x, long double *ip)
1619{
68611e6f
JH
1620 *ip = aintl(x);
1621 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
55954f19 1622}
d67dac15 1623# endif
55954f19
JH
1624#endif
1625
7b9b7dff 1626/* Similarly, with ilogbl and scalbnl we can emulate frexpl. */
55954f19
JH
1627#if ! defined(HAS_FREXPL) && defined(HAS_ILOGBL) && defined(HAS_SCALBNL)
1628long double
1629Perl_my_frexpl(long double x, int *e) {
68611e6f
JH
1630 *e = x == 0.0L ? 0 : ilogbl(x) + 1;
1631 return (scalbnl(x, -*e));
55954f19
JH
1632}
1633#endif
66610fdd
RGS
1634
1635/*
ed140128
AD
1636=for apidoc Perl_signbit
1637
1638Return a non-zero integer if the sign bit on an NV is set, and 0 if
1639it is not.
1640
796b6530
KW
1641If F<Configure> detects this system has a C<signbit()> that will work with
1642our NVs, then we just use it via the C<#define> in F<perl.h>. Otherwise,
8b7fad81 1643fall back on this implementation. The main use of this function
796b6530 1644is catching C<-0.0>.
ed140128 1645
796b6530
KW
1646C<Configure> notes: This function is called C<'Perl_signbit'> instead of a
1647plain C<'signbit'> because it is easy to imagine a system having a C<signbit()>
ed140128 1648function or macro that doesn't happen to work with our particular choice
796b6530 1649of NVs. We shouldn't just re-C<#define> C<signbit> as C<Perl_signbit> and expect
ed140128 1650the standard system headers to be happy. Also, this is a no-context
796b6530
KW
1651function (no C<pTHX_>) because C<Perl_signbit()> is usually re-C<#defined> in
1652F<perl.h> as a simple macro call to the system's C<signbit()>.
1653Users should just always call C<Perl_signbit()>.
ed140128
AD
1654
1655=cut
1656*/
1657#if !defined(HAS_SIGNBIT)
1658int
1659Perl_signbit(NV x) {
8b7fad81 1660# ifdef Perl_fp_class_nzero
406d5545
JH
1661 return Perl_fp_class_nzero(x);
1662 /* Try finding the high byte, and assume it's highest bit
1663 * is the sign. This assumption is probably wrong somewhere. */
572cd850
JH
1664# elif defined(USE_LONG_DOUBLE) && LONG_DOUBLEKIND == LONG_DOUBLE_IS_X86_80_BIT_LITTLE_ENDIAN
1665 return (((unsigned char *)&x)[9] & 0x80);
1666# elif defined(NV_LITTLE_ENDIAN)
1667 /* Note that NVSIZE is sizeof(NV), which would make the below be
1668 * wrong if the end bytes are unused, which happens with the x86
1669 * 80-bit long doubles, which is why take care of that above. */
1670 return (((unsigned char *)&x)[NVSIZE - 1] & 0x80);
1671# elif defined(NV_BIG_ENDIAN)
1672 return (((unsigned char *)&x)[0] & 0x80);
1673# else
406d5545 1674 /* This last resort fallback is wrong for the negative zero. */
3585840c 1675 return (x < 0.0) ? 1 : 0;
572cd850 1676# endif
ed140128
AD
1677}
1678#endif
1679
1680/*
14d04a33 1681 * ex: set ts=8 sts=4 sw=4 et:
37442d52 1682 */