This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
APItest/t/utf8.t: Fix improper tests
[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{
521#ifdef USE_LOCALE_NUMERIC
7918f24d
NC
522 PERL_ARGS_ASSERT_GROK_NUMERIC_RADIX;
523
d6ded950 524 if (IN_LC(LC_NUMERIC)) {
67d796ae
KW
525 DECLARATION_FOR_LC_NUMERIC_MANIPULATION;
526 STORE_LC_NUMERIC_SET_TO_NEEDED();
21431899 527 if (PL_numeric_radix_sv) {
c5a7e38e
KW
528 STRLEN len;
529 const char * const radix = SvPV(PL_numeric_radix_sv, len);
530 if (*sp + len <= send && memEQ(*sp, radix, len)) {
531 *sp += len;
532 RESTORE_LC_NUMERIC();
533 return TRUE;
534 }
21431899
KW
535 }
536 RESTORE_LC_NUMERIC();
98994639
HS
537 }
538 /* always try "." if numeric radix didn't match because
539 * we may have data from different locales mixed */
540#endif
7918f24d
NC
541
542 PERL_ARGS_ASSERT_GROK_NUMERIC_RADIX;
543
98994639
HS
544 if (*sp < send && **sp == '.') {
545 ++*sp;
546 return TRUE;
547 }
548 return FALSE;
549}
550
569f27e5 551/*
ff4eb398
JH
552=for apidoc grok_infnan
553
796b6530 554Helper for C<grok_number()>, accepts various ways of spelling "infinity"
ff4eb398
JH
555or "not a number", and returns one of the following flag combinations:
556
557 IS_NUMBER_INFINITE
558 IS_NUMBER_NAN
559 IS_NUMBER_INFINITE | IS_NUMBER_NEG
560 IS_NUMBER_NAN | IS_NUMBER_NEG
561 0
562
796b6530 563possibly |-ed with C<IS_NUMBER_TRAILING>.
b489e20f 564
796b6530 565If an infinity or a not-a-number is recognized, C<*sp> will point to
62bdc035 566one byte past the end of the recognized string. If the recognition fails,
796b6530 567zero is returned, and C<*sp> will not move.
ff4eb398
JH
568
569=cut
570*/
571
572int
3823048b 573Perl_grok_infnan(pTHX_ const char** sp, const char* send)
ff4eb398
JH
574{
575 const char* s = *sp;
576 int flags = 0;
a5dc2484 577#if defined(NV_INF) || defined(NV_NAN)
62bdc035 578 bool odh = FALSE; /* one-dot-hash: 1.#INF */
ff4eb398
JH
579
580 PERL_ARGS_ASSERT_GROK_INFNAN;
581
8c12dc63
JH
582 if (*s == '+') {
583 s++; if (s == send) return 0;
584 }
585 else if (*s == '-') {
ff4eb398
JH
586 flags |= IS_NUMBER_NEG; /* Yes, -NaN happens. Incorrect but happens. */
587 s++; if (s == send) return 0;
588 }
589
590 if (*s == '1') {
62bdc035
JH
591 /* Visual C: 1.#SNAN, -1.#QNAN, 1#INF, 1.#IND (maybe also 1.#NAN)
592 * Let's keep the dot optional. */
ff4eb398
JH
593 s++; if (s == send) return 0;
594 if (*s == '.') {
595 s++; if (s == send) return 0;
596 }
597 if (*s == '#') {
598 s++; if (s == send) return 0;
599 } else
600 return 0;
e855f543 601 odh = TRUE;
ff4eb398
JH
602 }
603
305b8651 604 if (isALPHA_FOLD_EQ(*s, 'I')) {
62bdc035
JH
605 /* INF or IND (1.#IND is "indeterminate", a certain type of NAN) */
606
305b8651 607 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
ff4eb398 608 s++; if (s == send) return 0;
305b8651 609 if (isALPHA_FOLD_EQ(*s, 'F')) {
ff4eb398 610 s++;
b8974fcb
JH
611 if (s < send && (isALPHA_FOLD_EQ(*s, 'I'))) {
612 int fail =
613 flags | IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT | IS_NUMBER_TRAILING;
614 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return fail;
615 s++; if (s == send || isALPHA_FOLD_NE(*s, 'I')) return fail;
616 s++; if (s == send || isALPHA_FOLD_NE(*s, 'T')) return fail;
617 s++; if (s == send || isALPHA_FOLD_NE(*s, 'Y')) return fail;
3396ed30 618 s++;
b8974fcb
JH
619 } else if (odh) {
620 while (*s == '0') { /* 1.#INF00 */
621 s++;
622 }
3396ed30 623 }
b489e20f
JH
624 while (s < send && isSPACE(*s))
625 s++;
626 if (s < send && *s) {
3396ed30 627 flags |= IS_NUMBER_TRAILING;
fae4db12 628 }
ff4eb398
JH
629 flags |= IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT;
630 }
e855f543 631 else if (isALPHA_FOLD_EQ(*s, 'D') && odh) { /* 1.#IND */
ff4eb398
JH
632 s++;
633 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
fae4db12
JH
634 while (*s == '0') { /* 1.#IND00 */
635 s++;
636 }
1e9aa12f
JH
637 if (*s) {
638 flags |= IS_NUMBER_TRAILING;
639 }
ff4eb398
JH
640 } else
641 return 0;
ff4eb398
JH
642 }
643 else {
62bdc035 644 /* Maybe NAN of some sort */
3823048b
JH
645
646 if (isALPHA_FOLD_EQ(*s, 'S') || isALPHA_FOLD_EQ(*s, 'Q')) {
647 /* snan, qNaN */
648 /* XXX do something with the snan/qnan difference */
649 s++; if (s == send) return 0;
650 }
651
652 if (isALPHA_FOLD_EQ(*s, 'N')) {
653 s++; if (s == send || isALPHA_FOLD_NE(*s, 'A')) return 0;
654 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
655 s++;
656
657 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
658
659 /* NaN can be followed by various stuff (NaNQ, NaNS), but
660 * there are also multiple different NaN values, and some
661 * implementations output the "payload" values,
662 * e.g. NaN123, NAN(abc), while some legacy implementations
663 * have weird stuff like NaN%. */
664 if (isALPHA_FOLD_EQ(*s, 'q') ||
665 isALPHA_FOLD_EQ(*s, 's')) {
666 /* "nanq" or "nans" are ok, though generating
667 * these portably is tricky. */
668 s++;
669 }
670 if (*s == '(') {
671 /* C99 style "nan(123)" or Perlish equivalent "nan($uv)". */
672 const char *t;
673 s++;
674 if (s == send) {
675 return flags | IS_NUMBER_TRAILING;
676 }
677 t = s + 1;
678 while (t < send && *t && *t != ')') {
679 t++;
680 }
681 if (t == send) {
682 return flags | IS_NUMBER_TRAILING;
683 }
684 if (*t == ')') {
685 int nantype;
686 UV nanval;
687 if (s[0] == '0' && s + 2 < t &&
688 isALPHA_FOLD_EQ(s[1], 'x') &&
689 isXDIGIT(s[2])) {
690 STRLEN len = t - s;
691 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES;
692 nanval = grok_hex(s, &len, &flags, NULL);
693 if ((flags & PERL_SCAN_GREATER_THAN_UV_MAX)) {
694 nantype = 0;
695 } else {
696 nantype = IS_NUMBER_IN_UV;
697 }
698 s += len;
699 } else if (s[0] == '0' && s + 2 < t &&
700 isALPHA_FOLD_EQ(s[1], 'b') &&
701 (s[2] == '0' || s[2] == '1')) {
702 STRLEN len = t - s;
703 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES;
704 nanval = grok_bin(s, &len, &flags, NULL);
705 if ((flags & PERL_SCAN_GREATER_THAN_UV_MAX)) {
706 nantype = 0;
707 } else {
708 nantype = IS_NUMBER_IN_UV;
709 }
710 s += len;
711 } else {
712 const char *u;
713 nantype =
714 grok_number_flags(s, t - s, &nanval,
715 PERL_SCAN_TRAILING |
716 PERL_SCAN_ALLOW_UNDERSCORES);
717 /* Unfortunately grok_number_flags() doesn't
718 * tell how far we got and the ')' will always
719 * be "trailing", so we need to double-check
720 * whether we had something dubious. */
721 for (u = s; u < t; u++) {
722 if (!isDIGIT(*u)) {
723 flags |= IS_NUMBER_TRAILING;
724 break;
725 }
726 }
727 s = u;
728 }
729
730 /* XXX Doesn't do octal: nan("0123").
731 * Probably not a big loss. */
732
733 if ((nantype & IS_NUMBER_NOT_INT) ||
734 !(nantype && IS_NUMBER_IN_UV)) {
735 /* XXX the nanval is currently unused, that is,
736 * not inserted as the NaN payload of the NV.
737 * But the above code already parses the C99
738 * nan(...) format. See below, and see also
739 * the nan() in POSIX.xs.
740 *
741 * Certain configuration combinations where
742 * NVSIZE is greater than UVSIZE mean that
743 * a single UV cannot contain all the possible
744 * NaN payload bits. There would need to be
745 * some more generic syntax than "nan($uv)".
746 *
747 * Issues to keep in mind:
748 *
749 * (1) In most common cases there would
750 * not be an integral number of bytes that
751 * could be set, only a certain number of bits.
752 * For example for the common case of
753 * NVSIZE == UVSIZE == 8 there is room for 52
754 * bits in the payload, but the most significant
755 * bit is commonly reserved for the
756 * signaling/quiet bit, leaving 51 bits.
757 * Furthermore, the C99 nan() is supposed
758 * to generate quiet NaNs, so it is doubtful
759 * whether it should be able to generate
760 * signaling NaNs. For the x86 80-bit doubles
761 * (if building a long double Perl) there would
762 * be 62 bits (s/q bit being the 63rd).
763 *
764 * (2) Endianness of the payload bits. If the
765 * payload is specified as an UV, the low-order
766 * bits of the UV are naturally little-endianed
767 * (rightmost) bits of the payload. The endianness
768 * of UVs and NVs can be different. */
769 return 0;
770 }
771 if (s < t) {
772 flags |= IS_NUMBER_TRAILING;
773 }
774 } else {
775 /* Looked like nan(...), but no close paren. */
776 flags |= IS_NUMBER_TRAILING;
777 }
778 } else {
779 while (s < send && isSPACE(*s))
780 s++;
781 if (s < send && *s) {
782 /* Note that we here implicitly accept (parse as
783 * "nan", but with warnings) also any other weird
784 * trailing stuff for "nan". In the above we just
785 * check that if we got the C99-style "nan(...)",
786 * the "..." looks sane.
787 * If in future we accept more ways of specifying
788 * the nan payload, the accepting would happen around
789 * here. */
790 flags |= IS_NUMBER_TRAILING;
791 }
792 }
793 s = send;
794 }
795 else
796 return 0;
ff4eb398
JH
797 }
798
b489e20f
JH
799 while (s < send && isSPACE(*s))
800 s++;
801
a5dc2484
JH
802#else
803 PERL_UNUSED_ARG(send);
804#endif /* #if defined(NV_INF) || defined(NV_NAN) */
a1fe7cea
JH
805 *sp = s;
806 return flags;
ff4eb398
JH
807}
808
13393a5e 809/*
3823048b 810=for apidoc grok_number_flags
13393a5e
JH
811
812Recognise (or not) a number. The type of the number is returned
813(0 if unrecognised), otherwise it is a bit-ORed combination of
796b6530
KW
814C<IS_NUMBER_IN_UV>, C<IS_NUMBER_GREATER_THAN_UV_MAX>, C<IS_NUMBER_NOT_INT>,
815C<IS_NUMBER_NEG>, C<IS_NUMBER_INFINITY>, C<IS_NUMBER_NAN> (defined in perl.h).
816
817If the value of the number can fit in a UV, it is returned in C<*valuep>.
818C<IS_NUMBER_IN_UV> will be set to indicate that C<*valuep> is valid, C<IS_NUMBER_IN_UV>
819will never be set unless C<*valuep> is valid, but C<*valuep> may have been assigned
820to during processing even though C<IS_NUMBER_IN_UV> is not set on return.
821If C<valuep> is C<NULL>, C<IS_NUMBER_IN_UV> will be set for the same cases as when
822C<valuep> is non-C<NULL>, but no actual assignment (or SEGV) will occur.
823
824C<IS_NUMBER_NOT_INT> will be set with C<IS_NUMBER_IN_UV> if trailing decimals were
825seen (in which case C<*valuep> gives the true value truncated to an integer), and
826C<IS_NUMBER_NEG> if the number is negative (in which case C<*valuep> holds the
827absolute value). C<IS_NUMBER_IN_UV> is not set if e notation was used or the
13393a5e
JH
828number is larger than a UV.
829
830C<flags> allows only C<PERL_SCAN_TRAILING>, which allows for trailing
831non-numeric text on an otherwise successful I<grok>, setting
832C<IS_NUMBER_TRAILING> on the result.
833
834=for apidoc grok_number
835
796b6530 836Identical to C<grok_number_flags()> with C<flags> set to zero.
13393a5e
JH
837
838=cut
839 */
840int
841Perl_grok_number(pTHX_ const char *pv, STRLEN len, UV *valuep)
842{
843 PERL_ARGS_ASSERT_GROK_NUMBER;
844
845 return grok_number_flags(pv, len, valuep, 0);
846}
847
945b524a
JH
848static const UV uv_max_div_10 = UV_MAX / 10;
849static const U8 uv_max_mod_10 = UV_MAX % 10;
850
3f7602fa 851int
3823048b 852Perl_grok_number_flags(pTHX_ const char *pv, STRLEN len, UV *valuep, U32 flags)
3f7602fa 853{
60939fb8 854 const char *s = pv;
c4420975 855 const char * const send = pv + len;
ae776a2c 856 const char *d;
60939fb8 857 int numtype = 0;
60939fb8 858
3823048b 859 PERL_ARGS_ASSERT_GROK_NUMBER_FLAGS;
7918f24d 860
60939fb8
NC
861 while (s < send && isSPACE(*s))
862 s++;
863 if (s == send) {
864 return 0;
865 } else if (*s == '-') {
866 s++;
867 numtype = IS_NUMBER_NEG;
868 }
869 else if (*s == '+')
aa42a541 870 s++;
60939fb8
NC
871
872 if (s == send)
873 return 0;
874
ae776a2c 875 /* The first digit (after optional sign): note that might
8c12dc63 876 * also point to "infinity" or "nan", or "1.#INF". */
ae776a2c
JH
877 d = s;
878
8c12dc63 879 /* next must be digit or the radix separator or beginning of infinity/nan */
60939fb8
NC
880 if (isDIGIT(*s)) {
881 /* UVs are at least 32 bits, so the first 9 decimal digits cannot
882 overflow. */
883 UV value = *s - '0';
884 /* This construction seems to be more optimiser friendly.
885 (without it gcc does the isDIGIT test and the *s - '0' separately)
886 With it gcc on arm is managing 6 instructions (6 cycles) per digit.
887 In theory the optimiser could deduce how far to unroll the loop
888 before checking for overflow. */
58bb9ec3
NC
889 if (++s < send) {
890 int digit = *s - '0';
60939fb8
NC
891 if (digit >= 0 && digit <= 9) {
892 value = value * 10 + digit;
58bb9ec3
NC
893 if (++s < send) {
894 digit = *s - '0';
60939fb8
NC
895 if (digit >= 0 && digit <= 9) {
896 value = value * 10 + digit;
58bb9ec3
NC
897 if (++s < send) {
898 digit = *s - '0';
60939fb8
NC
899 if (digit >= 0 && digit <= 9) {
900 value = value * 10 + digit;
58bb9ec3
NC
901 if (++s < send) {
902 digit = *s - '0';
60939fb8
NC
903 if (digit >= 0 && digit <= 9) {
904 value = value * 10 + digit;
58bb9ec3
NC
905 if (++s < send) {
906 digit = *s - '0';
60939fb8
NC
907 if (digit >= 0 && digit <= 9) {
908 value = value * 10 + digit;
58bb9ec3
NC
909 if (++s < send) {
910 digit = *s - '0';
60939fb8
NC
911 if (digit >= 0 && digit <= 9) {
912 value = value * 10 + digit;
58bb9ec3
NC
913 if (++s < send) {
914 digit = *s - '0';
60939fb8
NC
915 if (digit >= 0 && digit <= 9) {
916 value = value * 10 + digit;
58bb9ec3
NC
917 if (++s < send) {
918 digit = *s - '0';
60939fb8
NC
919 if (digit >= 0 && digit <= 9) {
920 value = value * 10 + digit;
58bb9ec3 921 if (++s < send) {
60939fb8
NC
922 /* Now got 9 digits, so need to check
923 each time for overflow. */
58bb9ec3 924 digit = *s - '0';
60939fb8 925 while (digit >= 0 && digit <= 9
945b524a
JH
926 && (value < uv_max_div_10
927 || (value == uv_max_div_10
928 && digit <= uv_max_mod_10))) {
60939fb8 929 value = value * 10 + digit;
58bb9ec3
NC
930 if (++s < send)
931 digit = *s - '0';
60939fb8
NC
932 else
933 break;
934 }
935 if (digit >= 0 && digit <= 9
51bd16da 936 && (s < send)) {
60939fb8
NC
937 /* value overflowed.
938 skip the remaining digits, don't
939 worry about setting *valuep. */
940 do {
941 s++;
942 } while (s < send && isDIGIT(*s));
943 numtype |=
944 IS_NUMBER_GREATER_THAN_UV_MAX;
945 goto skip_value;
946 }
947 }
948 }
98994639 949 }
60939fb8
NC
950 }
951 }
952 }
953 }
954 }
955 }
956 }
957 }
958 }
959 }
960 }
98994639 961 }
60939fb8 962 }
98994639 963 }
60939fb8
NC
964 numtype |= IS_NUMBER_IN_UV;
965 if (valuep)
966 *valuep = value;
967
968 skip_value:
969 if (GROK_NUMERIC_RADIX(&s, send)) {
970 numtype |= IS_NUMBER_NOT_INT;
971 while (s < send && isDIGIT(*s)) /* optional digits after the radix */
972 s++;
98994639 973 }
60939fb8
NC
974 }
975 else if (GROK_NUMERIC_RADIX(&s, send)) {
976 numtype |= IS_NUMBER_NOT_INT | IS_NUMBER_IN_UV; /* valuep assigned below */
977 /* no digits before the radix means we need digits after it */
978 if (s < send && isDIGIT(*s)) {
979 do {
980 s++;
981 } while (s < send && isDIGIT(*s));
982 if (valuep) {
983 /* integer approximation is valid - it's 0. */
984 *valuep = 0;
985 }
98994639 986 }
60939fb8 987 else
ae776a2c 988 return 0;
ff4eb398 989 }
60939fb8 990
926f5fc6 991 if (s > d && s < send) {
60939fb8 992 /* we can have an optional exponent part */
305b8651 993 if (isALPHA_FOLD_EQ(*s, 'e')) {
60939fb8
NC
994 s++;
995 if (s < send && (*s == '-' || *s == '+'))
996 s++;
997 if (s < send && isDIGIT(*s)) {
998 do {
999 s++;
1000 } while (s < send && isDIGIT(*s));
1001 }
3f7602fa
TC
1002 else if (flags & PERL_SCAN_TRAILING)
1003 return numtype | IS_NUMBER_TRAILING;
60939fb8 1004 else
3f7602fa
TC
1005 return 0;
1006
1007 /* The only flag we keep is sign. Blow away any "it's UV" */
1008 numtype &= IS_NUMBER_NEG;
1009 numtype |= IS_NUMBER_NOT_INT;
60939fb8
NC
1010 }
1011 }
1012 while (s < send && isSPACE(*s))
1013 s++;
1014 if (s >= send)
aa8b85de 1015 return numtype;
60939fb8
NC
1016 if (len == 10 && memEQ(pv, "0 but true", 10)) {
1017 if (valuep)
1018 *valuep = 0;
1019 return IS_NUMBER_IN_UV;
1020 }
8c12dc63
JH
1021 /* We could be e.g. at "Inf" or "NaN", or at the "#" of "1.#INF". */
1022 if ((s + 2 < send) && strchr("inqs#", toFOLD(*s))) {
1023 /* Really detect inf/nan. Start at d, not s, since the above
1024 * code might have already consumed the "1." or "1". */
3823048b 1025 int infnan = Perl_grok_infnan(aTHX_ &d, send);
8c12dc63
JH
1026 if ((infnan & IS_NUMBER_INFINITY)) {
1027 return (numtype | infnan); /* Keep sign for infinity. */
1028 }
1029 else if ((infnan & IS_NUMBER_NAN)) {
1030 return (numtype | infnan) & ~IS_NUMBER_NEG; /* Clear sign for nan. */
1031 }
1032 }
3f7602fa
TC
1033 else if (flags & PERL_SCAN_TRAILING) {
1034 return numtype | IS_NUMBER_TRAILING;
1035 }
1036
60939fb8 1037 return 0;
98994639
HS
1038}
1039
6313e544 1040/*
73e43954 1041grok_atoUV
6313e544 1042
22ff3130 1043grok_atoUV parses a C-style zero-byte terminated string, looking for
d62b8c6a 1044a decimal unsigned integer.
338aa8b0 1045
d62b8c6a
JH
1046Returns the unsigned integer, if a valid value can be parsed
1047from the beginning of the string.
f4379102 1048
d62b8c6a 1049Accepts only the decimal digits '0'..'9'.
6313e544 1050
22ff3130 1051As opposed to atoi or strtol, grok_atoUV does NOT allow optional
d62b8c6a
JH
1052leading whitespace, or negative inputs. If such features are
1053required, the calling code needs to explicitly implement those.
6313e544 1054
22ff3130
HS
1055Returns true if a valid value could be parsed. In that case, valptr
1056is set to the parsed value, and endptr (if provided) is set to point
1057to the character after the last digit.
338aa8b0 1058
22ff3130
HS
1059Returns false otherwise. This can happen if a) there is a leading zero
1060followed by another digit; b) the digits would overflow a UV; or c)
1061there are trailing non-digits AND endptr is not provided.
6313e544 1062
d62b8c6a
JH
1063Background: atoi has severe problems with illegal inputs, it cannot be
1064used for incremental parsing, and therefore should be avoided
1065atoi and strtol are also affected by locale settings, which can also be
1066seen as a bug (global state controlled by user environment).
1067
6313e544
JH
1068*/
1069
22ff3130
HS
1070bool
1071Perl_grok_atoUV(const char *pv, UV *valptr, const char** endptr)
6313e544
JH
1072{
1073 const char* s = pv;
1074 const char** eptr;
1075 const char* end2; /* Used in case endptr is NULL. */
22ff3130 1076 UV val = 0; /* The parsed value. */
6313e544 1077
22ff3130 1078 PERL_ARGS_ASSERT_GROK_ATOUV;
6313e544
JH
1079
1080 eptr = endptr ? endptr : &end2;
75feedba
JH
1081 if (isDIGIT(*s)) {
1082 /* Single-digit inputs are quite common. */
6313e544 1083 val = *s++ - '0';
75feedba 1084 if (isDIGIT(*s)) {
22ff3130
HS
1085 /* Fail on extra leading zeros. */
1086 if (val == 0)
1087 return FALSE;
75feedba
JH
1088 while (isDIGIT(*s)) {
1089 /* This could be unrolled like in grok_number(), but
1090 * the expected uses of this are not speed-needy, and
1091 * unlikely to need full 64-bitness. */
1092 U8 digit = *s++ - '0';
945b524a
JH
1093 if (val < uv_max_div_10 ||
1094 (val == uv_max_div_10 && digit <= uv_max_mod_10)) {
75feedba
JH
1095 val = val * 10 + digit;
1096 } else {
22ff3130 1097 return FALSE;
6313e544 1098 }
6313e544
JH
1099 }
1100 }
75feedba 1101 }
22ff3130
HS
1102 if (s == pv)
1103 return FALSE;
1104 if (endptr == NULL && *s)
1105 return FALSE; /* If endptr is NULL, no trailing non-digits allowed. */
6313e544 1106 *eptr = s;
22ff3130
HS
1107 *valptr = val;
1108 return TRUE;
6313e544
JH
1109}
1110
a4eca1d4 1111#ifndef USE_QUADMATH
4801ca72 1112STATIC NV
98994639
HS
1113S_mulexp10(NV value, I32 exponent)
1114{
1115 NV result = 1.0;
1116 NV power = 10.0;
1117 bool negative = 0;
1118 I32 bit;
1119
1120 if (exponent == 0)
1121 return value;
659c4b96
DM
1122 if (value == 0)
1123 return (NV)0;
87032ba1 1124
24866caa 1125 /* On OpenVMS VAX we by default use the D_FLOAT double format,
67597c89 1126 * and that format does not have *easy* capabilities [1] for
24866caa
CB
1127 * overflowing doubles 'silently' as IEEE fp does. We also need
1128 * to support G_FLOAT on both VAX and Alpha, and though the exponent
1129 * range is much larger than D_FLOAT it still doesn't do silent
1130 * overflow. Therefore we need to detect early whether we would
1131 * overflow (this is the behaviour of the native string-to-float
1132 * conversion routines, and therefore of native applications, too).
67597c89 1133 *
24866caa
CB
1134 * [1] Trying to establish a condition handler to trap floating point
1135 * exceptions is not a good idea. */
87032ba1
JH
1136
1137 /* In UNICOS and in certain Cray models (such as T90) there is no
1138 * IEEE fp, and no way at all from C to catch fp overflows gracefully.
1139 * There is something you can do if you are willing to use some
1140 * inline assembler: the instruction is called DFI-- but that will
1141 * disable *all* floating point interrupts, a little bit too large
1142 * a hammer. Therefore we need to catch potential overflows before
1143 * it's too late. */
353813d9 1144
a7157111 1145#if ((defined(VMS) && !defined(_IEEE_FP)) || defined(_UNICOS) || defined(DOUBLE_IS_VAX_FLOAT)) && defined(NV_MAX_10_EXP)
353813d9 1146 STMT_START {
c4420975 1147 const NV exp_v = log10(value);
353813d9
HS
1148 if (exponent >= NV_MAX_10_EXP || exponent + exp_v >= NV_MAX_10_EXP)
1149 return NV_MAX;
1150 if (exponent < 0) {
1151 if (-(exponent + exp_v) >= NV_MAX_10_EXP)
1152 return 0.0;
1153 while (-exponent >= NV_MAX_10_EXP) {
1154 /* combination does not overflow, but 10^(-exponent) does */
1155 value /= 10;
1156 ++exponent;
1157 }
1158 }
1159 } STMT_END;
87032ba1
JH
1160#endif
1161
353813d9
HS
1162 if (exponent < 0) {
1163 negative = 1;
1164 exponent = -exponent;
b27804d8
DM
1165#ifdef NV_MAX_10_EXP
1166 /* for something like 1234 x 10^-309, the action of calculating
1167 * the intermediate value 10^309 then returning 1234 / (10^309)
1168 * will fail, since 10^309 becomes infinity. In this case try to
1169 * refactor it as 123 / (10^308) etc.
1170 */
1171 while (value && exponent > NV_MAX_10_EXP) {
1172 exponent--;
1173 value /= 10;
1174 }
48853916
JH
1175 if (value == 0.0)
1176 return value;
b27804d8 1177#endif
353813d9 1178 }
c62e754c
JH
1179#if defined(__osf__)
1180 /* Even with cc -ieee + ieee_set_fp_control(IEEE_TRAP_ENABLE_INV)
1181 * Tru64 fp behavior on inf/nan is somewhat broken. Another way
1182 * to do this would be ieee_set_fp_control(IEEE_TRAP_ENABLE_OVF)
1183 * but that breaks another set of infnan.t tests. */
1184# define FP_OVERFLOWS_TO_ZERO
1185#endif
98994639
HS
1186 for (bit = 1; exponent; bit <<= 1) {
1187 if (exponent & bit) {
1188 exponent ^= bit;
1189 result *= power;
c62e754c
JH
1190#ifdef FP_OVERFLOWS_TO_ZERO
1191 if (result == 0)
a7157111 1192# ifdef NV_INF
c62e754c 1193 return value < 0 ? -NV_INF : NV_INF;
a7157111
JH
1194# else
1195 return value < 0 ? -FLT_MAX : FLT_MAX;
1196# endif
c62e754c 1197#endif
236f0012
CB
1198 /* Floating point exceptions are supposed to be turned off,
1199 * but if we're obviously done, don't risk another iteration.
1200 */
1201 if (exponent == 0) break;
98994639
HS
1202 }
1203 power *= power;
1204 }
1205 return negative ? value / result : value * result;
1206}
a4eca1d4 1207#endif /* #ifndef USE_QUADMATH */
98994639
HS
1208
1209NV
1210Perl_my_atof(pTHX_ const char* s)
1211{
1212 NV x = 0.0;
a4eca1d4
JH
1213#ifdef USE_QUADMATH
1214 Perl_my_atof2(aTHX_ s, &x);
1215 return x;
1216#else
1217# ifdef USE_LOCALE_NUMERIC
7918f24d
NC
1218 PERL_ARGS_ASSERT_MY_ATOF;
1219
a2287a13 1220 {
67d796ae
KW
1221 DECLARATION_FOR_LC_NUMERIC_MANIPULATION;
1222 STORE_LC_NUMERIC_SET_TO_NEEDED();
d6ded950 1223 if (PL_numeric_radix_sv && IN_LC(LC_NUMERIC)) {
e4850248
KW
1224 const char *standard = NULL, *local = NULL;
1225 bool use_standard_radix;
98994639 1226
e4850248
KW
1227 /* Look through the string for the first thing that looks like a
1228 * decimal point: either the value in the current locale or the
1229 * standard fallback of '.'. The one which appears earliest in the
1230 * input string is the one that we should have atof look for. Note
1231 * that we have to determine this beforehand because on some
1232 * systems, Perl_atof2 is just a wrapper around the system's atof.
1233 * */
1234 standard = strchr(s, '.');
1235 local = strstr(s, SvPV_nolen(PL_numeric_radix_sv));
78787052 1236
e4850248 1237 use_standard_radix = standard && (!local || standard < local);
78787052 1238
e4850248
KW
1239 if (use_standard_radix)
1240 SET_NUMERIC_STANDARD();
78787052 1241
e4850248 1242 Perl_atof2(s, x);
78787052 1243
e4850248 1244 if (use_standard_radix)
67d796ae 1245 SET_NUMERIC_UNDERLYING();
e4850248
KW
1246 }
1247 else
1248 Perl_atof2(s, x);
a2287a13
KW
1249 RESTORE_LC_NUMERIC();
1250 }
a4eca1d4 1251# else
a36244b7 1252 Perl_atof2(s, x);
a4eca1d4 1253# endif
98994639
HS
1254#endif
1255 return x;
1256}
1257
a7157111 1258#if defined(NV_INF) || defined(NV_NAN)
3c81f0b3
DD
1259
1260#ifdef USING_MSVC6
1261# pragma warning(push)
1262# pragma warning(disable:4756;disable:4056)
1263#endif
829757a4 1264static char*
5563f457 1265S_my_atof_infnan(pTHX_ const char* s, bool negative, const char* send, NV* value)
829757a4
JH
1266{
1267 const char *p0 = negative ? s - 1 : s;
1268 const char *p = p0;
3823048b 1269 int infnan = grok_infnan(&p, send);
829757a4
JH
1270 if (infnan && p != p0) {
1271 /* If we can generate inf/nan directly, let's do so. */
1272#ifdef NV_INF
1273 if ((infnan & IS_NUMBER_INFINITY)) {
3823048b 1274 *value = (infnan & IS_NUMBER_NEG) ? -NV_INF: NV_INF;
829757a4
JH
1275 return (char*)p;
1276 }
1277#endif
1278#ifdef NV_NAN
1279 if ((infnan & IS_NUMBER_NAN)) {
3823048b 1280 *value = NV_NAN;
829757a4
JH
1281 return (char*)p;
1282 }
1283#endif
1284#ifdef Perl_strtod
68611e6f 1285 /* If still here, we didn't have either NV_INF or NV_NAN,
829757a4
JH
1286 * and can try falling back to native strtod/strtold.
1287 *
1288 * The native interface might not recognize all the possible
1289 * inf/nan strings Perl recognizes. What we can try
1290 * is to try faking the input. We will try inf/-inf/nan
1291 * as the most promising/portable input. */
1292 {
1293 const char* fake = NULL;
1294 char* endp;
1295 NV nv;
a7157111 1296#ifdef NV_INF
829757a4
JH
1297 if ((infnan & IS_NUMBER_INFINITY)) {
1298 fake = ((infnan & IS_NUMBER_NEG)) ? "-inf" : "inf";
1299 }
a7157111
JH
1300#endif
1301#ifdef NV_NAN
1302 if ((infnan & IS_NUMBER_NAN)) {
829757a4
JH
1303 fake = "nan";
1304 }
a7157111 1305#endif
829757a4
JH
1306 assert(fake);
1307 nv = Perl_strtod(fake, &endp);
1308 if (fake != endp) {
a7157111 1309#ifdef NV_INF
829757a4 1310 if ((infnan & IS_NUMBER_INFINITY)) {
a7157111 1311# ifdef Perl_isinf
829757a4
JH
1312 if (Perl_isinf(nv))
1313 *value = nv;
a7157111 1314# else
829757a4
JH
1315 /* last resort, may generate SIGFPE */
1316 *value = Perl_exp((NV)1e9);
1317 if ((infnan & IS_NUMBER_NEG))
1318 *value = -*value;
a7157111 1319# endif
829757a4
JH
1320 return (char*)p; /* p, not endp */
1321 }
a7157111
JH
1322#endif
1323#ifdef NV_NAN
1324 if ((infnan & IS_NUMBER_NAN)) {
1325# ifdef Perl_isnan
829757a4
JH
1326 if (Perl_isnan(nv))
1327 *value = nv;
a7157111 1328# else
829757a4
JH
1329 /* last resort, may generate SIGFPE */
1330 *value = Perl_log((NV)-1.0);
a7157111 1331# endif
829757a4 1332 return (char*)p; /* p, not endp */
a7157111 1333#endif
829757a4
JH
1334 }
1335 }
1336 }
1337#endif /* #ifdef Perl_strtod */
1338 }
1339 return NULL;
1340}
3c81f0b3
DD
1341#ifdef USING_MSVC6
1342# pragma warning(pop)
1343#endif
829757a4 1344
a7157111
JH
1345#endif /* if defined(NV_INF) || defined(NV_NAN) */
1346
98994639
HS
1347char*
1348Perl_my_atof2(pTHX_ const char* orig, NV* value)
1349{
e1ec3a88 1350 const char* s = orig;
a4eca1d4
JH
1351 NV result[3] = {0.0, 0.0, 0.0};
1352#if defined(USE_PERL_ATOF) || defined(USE_QUADMATH)
ae776a2c 1353 const char* send = s + strlen(orig); /* one past the last */
a4eca1d4
JH
1354 bool negative = 0;
1355#endif
1356#if defined(USE_PERL_ATOF) && !defined(USE_QUADMATH)
1357 UV accumulator[2] = {0,0}; /* before/after dp */
8194bf88 1358 bool seen_digit = 0;
20f6aaab
AS
1359 I32 exp_adjust[2] = {0,0};
1360 I32 exp_acc[2] = {-1, -1};
1361 /* the current exponent adjust for the accumulators */
98994639 1362 I32 exponent = 0;
8194bf88 1363 I32 seen_dp = 0;
20f6aaab
AS
1364 I32 digit = 0;
1365 I32 old_digit = 0;
8194bf88 1366 I32 sig_digits = 0; /* noof significant digits seen so far */
a4eca1d4 1367#endif
8194bf88 1368
a4eca1d4 1369#if defined(USE_PERL_ATOF) || defined(USE_QUADMATH)
7918f24d
NC
1370 PERL_ARGS_ASSERT_MY_ATOF2;
1371
a4eca1d4
JH
1372 /* leading whitespace */
1373 while (isSPACE(*s))
1374 ++s;
1375
1376 /* sign */
1377 switch (*s) {
1378 case '-':
1379 negative = 1;
1380 /* FALLTHROUGH */
1381 case '+':
1382 ++s;
1383 }
1384#endif
1385
1386#ifdef USE_QUADMATH
1387 {
1388 char* endp;
adc55e02 1389 if ((endp = S_my_atof_infnan(aTHX_ s, negative, send, value)))
a4eca1d4
JH
1390 return endp;
1391 result[2] = strtoflt128(s, &endp);
1392 if (s != endp) {
1393 *value = negative ? -result[2] : result[2];
1394 return endp;
1395 }
1396 return NULL;
1397 }
1398#elif defined(USE_PERL_ATOF)
1399
8194bf88
DM
1400/* There is no point in processing more significant digits
1401 * than the NV can hold. Note that NV_DIG is a lower-bound value,
1402 * while we need an upper-bound value. We add 2 to account for this;
1403 * since it will have been conservative on both the first and last digit.
1404 * For example a 32-bit mantissa with an exponent of 4 would have
1405 * exact values in the set
1406 * 4
1407 * 8
1408 * ..
1409 * 17179869172
1410 * 17179869176
1411 * 17179869180
1412 *
1413 * where for the purposes of calculating NV_DIG we would have to discount
1414 * both the first and last digit, since neither can hold all values from
1415 * 0..9; but for calculating the value we must examine those two digits.
1416 */
ffa277e5
AS
1417#ifdef MAX_SIG_DIG_PLUS
1418 /* It is not necessarily the case that adding 2 to NV_DIG gets all the
1419 possible digits in a NV, especially if NVs are not IEEE compliant
1420 (e.g., long doubles on IRIX) - Allen <allens@cpan.org> */
1421# define MAX_SIG_DIGITS (NV_DIG+MAX_SIG_DIG_PLUS)
1422#else
1423# define MAX_SIG_DIGITS (NV_DIG+2)
1424#endif
8194bf88
DM
1425
1426/* the max number we can accumulate in a UV, and still safely do 10*N+9 */
1427#define MAX_ACCUMULATE ( (UV) ((UV_MAX - 9)/10))
98994639 1428
a5dc2484 1429#if defined(NV_INF) || defined(NV_NAN)
ae776a2c 1430 {
829757a4 1431 const char* endp;
5563f457 1432 if ((endp = S_my_atof_infnan(aTHX_ s, negative, send, value)))
829757a4 1433 return (char*)endp;
ae776a2c 1434 }
a5dc2484 1435#endif
2b54f59f 1436
8194bf88
DM
1437 /* we accumulate digits into an integer; when this becomes too
1438 * large, we add the total to NV and start again */
98994639 1439
8194bf88
DM
1440 while (1) {
1441 if (isDIGIT(*s)) {
1442 seen_digit = 1;
20f6aaab 1443 old_digit = digit;
8194bf88 1444 digit = *s++ - '0';
20f6aaab
AS
1445 if (seen_dp)
1446 exp_adjust[1]++;
98994639 1447
8194bf88
DM
1448 /* don't start counting until we see the first significant
1449 * digit, eg the 5 in 0.00005... */
1450 if (!sig_digits && digit == 0)
1451 continue;
1452
1453 if (++sig_digits > MAX_SIG_DIGITS) {
98994639 1454 /* limits of precision reached */
20f6aaab
AS
1455 if (digit > 5) {
1456 ++accumulator[seen_dp];
1457 } else if (digit == 5) {
1458 if (old_digit % 2) { /* round to even - Allen */
1459 ++accumulator[seen_dp];
1460 }
1461 }
1462 if (seen_dp) {
1463 exp_adjust[1]--;
1464 } else {
1465 exp_adjust[0]++;
1466 }
8194bf88 1467 /* skip remaining digits */
98994639 1468 while (isDIGIT(*s)) {
98994639 1469 ++s;
20f6aaab
AS
1470 if (! seen_dp) {
1471 exp_adjust[0]++;
1472 }
98994639
HS
1473 }
1474 /* warn of loss of precision? */
98994639 1475 }
8194bf88 1476 else {
20f6aaab 1477 if (accumulator[seen_dp] > MAX_ACCUMULATE) {
8194bf88 1478 /* add accumulator to result and start again */
20f6aaab
AS
1479 result[seen_dp] = S_mulexp10(result[seen_dp],
1480 exp_acc[seen_dp])
1481 + (NV)accumulator[seen_dp];
1482 accumulator[seen_dp] = 0;
1483 exp_acc[seen_dp] = 0;
98994639 1484 }
20f6aaab
AS
1485 accumulator[seen_dp] = accumulator[seen_dp] * 10 + digit;
1486 ++exp_acc[seen_dp];
98994639 1487 }
8194bf88 1488 }
e1ec3a88 1489 else if (!seen_dp && GROK_NUMERIC_RADIX(&s, send)) {
8194bf88 1490 seen_dp = 1;
20f6aaab 1491 if (sig_digits > MAX_SIG_DIGITS) {
c86f7df5 1492 do {
20f6aaab 1493 ++s;
c86f7df5 1494 } while (isDIGIT(*s));
20f6aaab
AS
1495 break;
1496 }
8194bf88
DM
1497 }
1498 else {
1499 break;
98994639
HS
1500 }
1501 }
1502
20f6aaab
AS
1503 result[0] = S_mulexp10(result[0], exp_acc[0]) + (NV)accumulator[0];
1504 if (seen_dp) {
1505 result[1] = S_mulexp10(result[1], exp_acc[1]) + (NV)accumulator[1];
1506 }
98994639 1507
305b8651 1508 if (seen_digit && (isALPHA_FOLD_EQ(*s, 'e'))) {
98994639
HS
1509 bool expnegative = 0;
1510
1511 ++s;
1512 switch (*s) {
1513 case '-':
1514 expnegative = 1;
924ba076 1515 /* FALLTHROUGH */
98994639
HS
1516 case '+':
1517 ++s;
1518 }
1519 while (isDIGIT(*s))
1520 exponent = exponent * 10 + (*s++ - '0');
1521 if (expnegative)
1522 exponent = -exponent;
1523 }
1524
20f6aaab
AS
1525
1526
98994639 1527 /* now apply the exponent */
20f6aaab
AS
1528
1529 if (seen_dp) {
1530 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0])
1531 + S_mulexp10(result[1],exponent-exp_adjust[1]);
1532 } else {
1533 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0]);
1534 }
98994639
HS
1535
1536 /* now apply the sign */
1537 if (negative)
20f6aaab 1538 result[2] = -result[2];
a36244b7 1539#endif /* USE_PERL_ATOF */
20f6aaab 1540 *value = result[2];
73d840c0 1541 return (char *)s;
98994639
HS
1542}
1543
5d34af89 1544/*
3d9d9213 1545=for apidoc isinfnan
5d34af89 1546
796b6530
KW
1547C<Perl_isinfnan()> is utility function that returns true if the NV
1548argument is either an infinity or a C<NaN>, false otherwise. To test
1549in more detail, use C<Perl_isinf()> and C<Perl_isnan()>.
5d34af89 1550
68611e6f
JH
1551This is also the logical inverse of Perl_isfinite().
1552
5d34af89
JH
1553=cut
1554*/
1cd88304
JH
1555bool
1556Perl_isinfnan(NV nv)
1557{
a5dc2484 1558 PERL_UNUSED_ARG(nv);
1cd88304
JH
1559#ifdef Perl_isinf
1560 if (Perl_isinf(nv))
1561 return TRUE;
1562#endif
1563#ifdef Perl_isnan
1564 if (Perl_isnan(nv))
1565 return TRUE;
1566#endif
1567 return FALSE;
1568}
1569
354b74ae
FC
1570/*
1571=for apidoc
1572
796b6530 1573Checks whether the argument would be either an infinity or C<NaN> when used
354b74ae 1574as a number, but is careful not to trigger non-numeric or uninitialized
796b6530 1575warnings. it assumes the caller has done C<SvGETMAGIC(sv)> already.
354b74ae
FC
1576
1577=cut
1578*/
1579
1580bool
1581Perl_isinfnansv(pTHX_ SV *sv)
1582{
1583 PERL_ARGS_ASSERT_ISINFNANSV;
1584 if (!SvOK(sv))
1585 return FALSE;
1586 if (SvNOKp(sv))
1587 return Perl_isinfnan(SvNVX(sv));
1588 if (SvIOKp(sv))
1589 return FALSE;
1590 {
1591 STRLEN len;
1592 const char *s = SvPV_nomg_const(sv, len);
3823048b 1593 return cBOOL(grok_infnan(&s, s+len));
354b74ae
FC
1594 }
1595}
1596
d67dac15 1597#ifndef HAS_MODFL
68611e6f
JH
1598/* C99 has truncl, pre-C99 Solaris had aintl. We can use either with
1599 * copysignl to emulate modfl, which is in some platforms missing or
1600 * broken. */
d67dac15
JH
1601# if defined(HAS_TRUNCL) && defined(HAS_COPYSIGNL)
1602long double
1603Perl_my_modfl(long double x, long double *ip)
1604{
68611e6f
JH
1605 *ip = truncl(x);
1606 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
d67dac15
JH
1607}
1608# elif defined(HAS_AINTL) && defined(HAS_COPYSIGNL)
55954f19
JH
1609long double
1610Perl_my_modfl(long double x, long double *ip)
1611{
68611e6f
JH
1612 *ip = aintl(x);
1613 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
55954f19 1614}
d67dac15 1615# endif
55954f19
JH
1616#endif
1617
7b9b7dff 1618/* Similarly, with ilogbl and scalbnl we can emulate frexpl. */
55954f19
JH
1619#if ! defined(HAS_FREXPL) && defined(HAS_ILOGBL) && defined(HAS_SCALBNL)
1620long double
1621Perl_my_frexpl(long double x, int *e) {
68611e6f
JH
1622 *e = x == 0.0L ? 0 : ilogbl(x) + 1;
1623 return (scalbnl(x, -*e));
55954f19
JH
1624}
1625#endif
66610fdd
RGS
1626
1627/*
ed140128
AD
1628=for apidoc Perl_signbit
1629
1630Return a non-zero integer if the sign bit on an NV is set, and 0 if
1631it is not.
1632
796b6530
KW
1633If F<Configure> detects this system has a C<signbit()> that will work with
1634our NVs, then we just use it via the C<#define> in F<perl.h>. Otherwise,
8b7fad81 1635fall back on this implementation. The main use of this function
796b6530 1636is catching C<-0.0>.
ed140128 1637
796b6530
KW
1638C<Configure> notes: This function is called C<'Perl_signbit'> instead of a
1639plain C<'signbit'> because it is easy to imagine a system having a C<signbit()>
ed140128 1640function or macro that doesn't happen to work with our particular choice
796b6530 1641of NVs. We shouldn't just re-C<#define> C<signbit> as C<Perl_signbit> and expect
ed140128 1642the standard system headers to be happy. Also, this is a no-context
796b6530
KW
1643function (no C<pTHX_>) because C<Perl_signbit()> is usually re-C<#defined> in
1644F<perl.h> as a simple macro call to the system's C<signbit()>.
1645Users should just always call C<Perl_signbit()>.
ed140128
AD
1646
1647=cut
1648*/
1649#if !defined(HAS_SIGNBIT)
1650int
1651Perl_signbit(NV x) {
8b7fad81 1652# ifdef Perl_fp_class_nzero
406d5545
JH
1653 return Perl_fp_class_nzero(x);
1654 /* Try finding the high byte, and assume it's highest bit
1655 * is the sign. This assumption is probably wrong somewhere. */
572cd850
JH
1656# elif defined(USE_LONG_DOUBLE) && LONG_DOUBLEKIND == LONG_DOUBLE_IS_X86_80_BIT_LITTLE_ENDIAN
1657 return (((unsigned char *)&x)[9] & 0x80);
1658# elif defined(NV_LITTLE_ENDIAN)
1659 /* Note that NVSIZE is sizeof(NV), which would make the below be
1660 * wrong if the end bytes are unused, which happens with the x86
1661 * 80-bit long doubles, which is why take care of that above. */
1662 return (((unsigned char *)&x)[NVSIZE - 1] & 0x80);
1663# elif defined(NV_BIG_ENDIAN)
1664 return (((unsigned char *)&x)[0] & 0x80);
1665# else
406d5545 1666 /* This last resort fallback is wrong for the negative zero. */
3585840c 1667 return (x < 0.0) ? 1 : 0;
572cd850 1668# endif
ed140128
AD
1669}
1670#endif
1671
1672/*
14d04a33 1673 * ex: set ts=8 sts=4 sw=4 et:
37442d52 1674 */