This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
embed.fnc: Fix my_atof2() entry
[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)) {
f0dafd73
KW
526 STRLEN len;
527 char * radix;
528 bool matches_radix = FALSE;
67d796ae 529 DECLARATION_FOR_LC_NUMERIC_MANIPULATION;
f0dafd73 530
a1395eaf 531 STORE_LC_NUMERIC_FORCE_TO_UNDERLYING();
f0dafd73
KW
532
533 radix = SvPV(PL_numeric_radix_sv, len);
534 radix = savepvn(radix, len);
535
21431899 536 RESTORE_LC_NUMERIC();
f0dafd73
KW
537
538 if (*sp + len <= send) {
539 matches_radix = memEQ(*sp, radix, len);
540 }
541
542 Safefree(radix);
543
544 if (matches_radix) {
545 *sp += len;
546 return TRUE;
547 }
98994639 548 }
f0dafd73 549
98994639 550#endif
7918f24d 551
f0dafd73
KW
552 /* always try "." if numeric radix didn't match because
553 * we may have data from different locales mixed */
98994639
HS
554 if (*sp < send && **sp == '.') {
555 ++*sp;
556 return TRUE;
557 }
f0dafd73 558
98994639
HS
559 return FALSE;
560}
561
569f27e5 562/*
ff4eb398
JH
563=for apidoc grok_infnan
564
796b6530 565Helper for C<grok_number()>, accepts various ways of spelling "infinity"
ff4eb398
JH
566or "not a number", and returns one of the following flag combinations:
567
568 IS_NUMBER_INFINITE
569 IS_NUMBER_NAN
570 IS_NUMBER_INFINITE | IS_NUMBER_NEG
571 IS_NUMBER_NAN | IS_NUMBER_NEG
572 0
573
796b6530 574possibly |-ed with C<IS_NUMBER_TRAILING>.
b489e20f 575
796b6530 576If an infinity or a not-a-number is recognized, C<*sp> will point to
62bdc035 577one byte past the end of the recognized string. If the recognition fails,
796b6530 578zero is returned, and C<*sp> will not move.
ff4eb398
JH
579
580=cut
581*/
582
583int
3823048b 584Perl_grok_infnan(pTHX_ const char** sp, const char* send)
ff4eb398
JH
585{
586 const char* s = *sp;
587 int flags = 0;
a5dc2484 588#if defined(NV_INF) || defined(NV_NAN)
62bdc035 589 bool odh = FALSE; /* one-dot-hash: 1.#INF */
ff4eb398
JH
590
591 PERL_ARGS_ASSERT_GROK_INFNAN;
592
8c12dc63
JH
593 if (*s == '+') {
594 s++; if (s == send) return 0;
595 }
596 else if (*s == '-') {
ff4eb398
JH
597 flags |= IS_NUMBER_NEG; /* Yes, -NaN happens. Incorrect but happens. */
598 s++; if (s == send) return 0;
599 }
600
601 if (*s == '1') {
62bdc035
JH
602 /* Visual C: 1.#SNAN, -1.#QNAN, 1#INF, 1.#IND (maybe also 1.#NAN)
603 * Let's keep the dot optional. */
ff4eb398
JH
604 s++; if (s == send) return 0;
605 if (*s == '.') {
606 s++; if (s == send) return 0;
607 }
608 if (*s == '#') {
609 s++; if (s == send) return 0;
610 } else
611 return 0;
e855f543 612 odh = TRUE;
ff4eb398
JH
613 }
614
305b8651 615 if (isALPHA_FOLD_EQ(*s, 'I')) {
62bdc035
JH
616 /* INF or IND (1.#IND is "indeterminate", a certain type of NAN) */
617
305b8651 618 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
ff4eb398 619 s++; if (s == send) return 0;
305b8651 620 if (isALPHA_FOLD_EQ(*s, 'F')) {
ff4eb398 621 s++;
b8974fcb
JH
622 if (s < send && (isALPHA_FOLD_EQ(*s, 'I'))) {
623 int fail =
624 flags | IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT | IS_NUMBER_TRAILING;
625 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return fail;
626 s++; if (s == send || isALPHA_FOLD_NE(*s, 'I')) return fail;
627 s++; if (s == send || isALPHA_FOLD_NE(*s, 'T')) return fail;
628 s++; if (s == send || isALPHA_FOLD_NE(*s, 'Y')) return fail;
3396ed30 629 s++;
b8974fcb
JH
630 } else if (odh) {
631 while (*s == '0') { /* 1.#INF00 */
632 s++;
633 }
3396ed30 634 }
b489e20f
JH
635 while (s < send && isSPACE(*s))
636 s++;
637 if (s < send && *s) {
3396ed30 638 flags |= IS_NUMBER_TRAILING;
fae4db12 639 }
ff4eb398
JH
640 flags |= IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT;
641 }
e855f543 642 else if (isALPHA_FOLD_EQ(*s, 'D') && odh) { /* 1.#IND */
ff4eb398
JH
643 s++;
644 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
fae4db12
JH
645 while (*s == '0') { /* 1.#IND00 */
646 s++;
647 }
1e9aa12f
JH
648 if (*s) {
649 flags |= IS_NUMBER_TRAILING;
650 }
ff4eb398
JH
651 } else
652 return 0;
ff4eb398
JH
653 }
654 else {
62bdc035 655 /* Maybe NAN of some sort */
3823048b
JH
656
657 if (isALPHA_FOLD_EQ(*s, 'S') || isALPHA_FOLD_EQ(*s, 'Q')) {
658 /* snan, qNaN */
659 /* XXX do something with the snan/qnan difference */
660 s++; if (s == send) return 0;
661 }
662
663 if (isALPHA_FOLD_EQ(*s, 'N')) {
664 s++; if (s == send || isALPHA_FOLD_NE(*s, 'A')) return 0;
665 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
666 s++;
667
668 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
669
670 /* NaN can be followed by various stuff (NaNQ, NaNS), but
671 * there are also multiple different NaN values, and some
672 * implementations output the "payload" values,
673 * e.g. NaN123, NAN(abc), while some legacy implementations
674 * have weird stuff like NaN%. */
675 if (isALPHA_FOLD_EQ(*s, 'q') ||
676 isALPHA_FOLD_EQ(*s, 's')) {
677 /* "nanq" or "nans" are ok, though generating
678 * these portably is tricky. */
679 s++;
680 }
681 if (*s == '(') {
682 /* C99 style "nan(123)" or Perlish equivalent "nan($uv)". */
683 const char *t;
684 s++;
685 if (s == send) {
686 return flags | IS_NUMBER_TRAILING;
687 }
688 t = s + 1;
689 while (t < send && *t && *t != ')') {
690 t++;
691 }
692 if (t == send) {
693 return flags | IS_NUMBER_TRAILING;
694 }
695 if (*t == ')') {
696 int nantype;
697 UV nanval;
698 if (s[0] == '0' && s + 2 < t &&
699 isALPHA_FOLD_EQ(s[1], 'x') &&
700 isXDIGIT(s[2])) {
701 STRLEN len = t - s;
702 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES;
703 nanval = grok_hex(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 if (s[0] == '0' && s + 2 < t &&
711 isALPHA_FOLD_EQ(s[1], 'b') &&
712 (s[2] == '0' || s[2] == '1')) {
713 STRLEN len = t - s;
714 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES;
715 nanval = grok_bin(s, &len, &flags, NULL);
716 if ((flags & PERL_SCAN_GREATER_THAN_UV_MAX)) {
717 nantype = 0;
718 } else {
719 nantype = IS_NUMBER_IN_UV;
720 }
721 s += len;
722 } else {
723 const char *u;
724 nantype =
725 grok_number_flags(s, t - s, &nanval,
726 PERL_SCAN_TRAILING |
727 PERL_SCAN_ALLOW_UNDERSCORES);
728 /* Unfortunately grok_number_flags() doesn't
729 * tell how far we got and the ')' will always
730 * be "trailing", so we need to double-check
731 * whether we had something dubious. */
732 for (u = s; u < t; u++) {
733 if (!isDIGIT(*u)) {
734 flags |= IS_NUMBER_TRAILING;
735 break;
736 }
737 }
738 s = u;
739 }
740
741 /* XXX Doesn't do octal: nan("0123").
742 * Probably not a big loss. */
743
744 if ((nantype & IS_NUMBER_NOT_INT) ||
745 !(nantype && IS_NUMBER_IN_UV)) {
746 /* XXX the nanval is currently unused, that is,
747 * not inserted as the NaN payload of the NV.
748 * But the above code already parses the C99
749 * nan(...) format. See below, and see also
750 * the nan() in POSIX.xs.
751 *
752 * Certain configuration combinations where
753 * NVSIZE is greater than UVSIZE mean that
754 * a single UV cannot contain all the possible
755 * NaN payload bits. There would need to be
756 * some more generic syntax than "nan($uv)".
757 *
758 * Issues to keep in mind:
759 *
760 * (1) In most common cases there would
761 * not be an integral number of bytes that
762 * could be set, only a certain number of bits.
763 * For example for the common case of
764 * NVSIZE == UVSIZE == 8 there is room for 52
765 * bits in the payload, but the most significant
766 * bit is commonly reserved for the
767 * signaling/quiet bit, leaving 51 bits.
768 * Furthermore, the C99 nan() is supposed
769 * to generate quiet NaNs, so it is doubtful
770 * whether it should be able to generate
771 * signaling NaNs. For the x86 80-bit doubles
772 * (if building a long double Perl) there would
773 * be 62 bits (s/q bit being the 63rd).
774 *
775 * (2) Endianness of the payload bits. If the
776 * payload is specified as an UV, the low-order
777 * bits of the UV are naturally little-endianed
778 * (rightmost) bits of the payload. The endianness
779 * of UVs and NVs can be different. */
780 return 0;
781 }
782 if (s < t) {
783 flags |= IS_NUMBER_TRAILING;
784 }
785 } else {
786 /* Looked like nan(...), but no close paren. */
787 flags |= IS_NUMBER_TRAILING;
788 }
789 } else {
790 while (s < send && isSPACE(*s))
791 s++;
792 if (s < send && *s) {
793 /* Note that we here implicitly accept (parse as
794 * "nan", but with warnings) also any other weird
795 * trailing stuff for "nan". In the above we just
796 * check that if we got the C99-style "nan(...)",
797 * the "..." looks sane.
798 * If in future we accept more ways of specifying
799 * the nan payload, the accepting would happen around
800 * here. */
801 flags |= IS_NUMBER_TRAILING;
802 }
803 }
804 s = send;
805 }
806 else
807 return 0;
ff4eb398
JH
808 }
809
b489e20f
JH
810 while (s < send && isSPACE(*s))
811 s++;
812
a5dc2484
JH
813#else
814 PERL_UNUSED_ARG(send);
815#endif /* #if defined(NV_INF) || defined(NV_NAN) */
a1fe7cea
JH
816 *sp = s;
817 return flags;
ff4eb398
JH
818}
819
13393a5e 820/*
3823048b 821=for apidoc grok_number_flags
13393a5e
JH
822
823Recognise (or not) a number. The type of the number is returned
824(0 if unrecognised), otherwise it is a bit-ORed combination of
796b6530
KW
825C<IS_NUMBER_IN_UV>, C<IS_NUMBER_GREATER_THAN_UV_MAX>, C<IS_NUMBER_NOT_INT>,
826C<IS_NUMBER_NEG>, C<IS_NUMBER_INFINITY>, C<IS_NUMBER_NAN> (defined in perl.h).
827
828If the value of the number can fit in a UV, it is returned in C<*valuep>.
829C<IS_NUMBER_IN_UV> will be set to indicate that C<*valuep> is valid, C<IS_NUMBER_IN_UV>
830will never be set unless C<*valuep> is valid, but C<*valuep> may have been assigned
831to during processing even though C<IS_NUMBER_IN_UV> is not set on return.
832If C<valuep> is C<NULL>, C<IS_NUMBER_IN_UV> will be set for the same cases as when
833C<valuep> is non-C<NULL>, but no actual assignment (or SEGV) will occur.
834
835C<IS_NUMBER_NOT_INT> will be set with C<IS_NUMBER_IN_UV> if trailing decimals were
836seen (in which case C<*valuep> gives the true value truncated to an integer), and
837C<IS_NUMBER_NEG> if the number is negative (in which case C<*valuep> holds the
838absolute value). C<IS_NUMBER_IN_UV> is not set if e notation was used or the
13393a5e
JH
839number is larger than a UV.
840
841C<flags> allows only C<PERL_SCAN_TRAILING>, which allows for trailing
842non-numeric text on an otherwise successful I<grok>, setting
843C<IS_NUMBER_TRAILING> on the result.
844
845=for apidoc grok_number
846
796b6530 847Identical to C<grok_number_flags()> with C<flags> set to zero.
13393a5e
JH
848
849=cut
850 */
851int
852Perl_grok_number(pTHX_ const char *pv, STRLEN len, UV *valuep)
853{
854 PERL_ARGS_ASSERT_GROK_NUMBER;
855
856 return grok_number_flags(pv, len, valuep, 0);
857}
858
945b524a
JH
859static const UV uv_max_div_10 = UV_MAX / 10;
860static const U8 uv_max_mod_10 = UV_MAX % 10;
861
3f7602fa 862int
3823048b 863Perl_grok_number_flags(pTHX_ const char *pv, STRLEN len, UV *valuep, U32 flags)
3f7602fa 864{
60939fb8 865 const char *s = pv;
c4420975 866 const char * const send = pv + len;
ae776a2c 867 const char *d;
60939fb8 868 int numtype = 0;
60939fb8 869
3823048b 870 PERL_ARGS_ASSERT_GROK_NUMBER_FLAGS;
7918f24d 871
60939fb8
NC
872 while (s < send && isSPACE(*s))
873 s++;
874 if (s == send) {
875 return 0;
876 } else if (*s == '-') {
877 s++;
878 numtype = IS_NUMBER_NEG;
879 }
880 else if (*s == '+')
aa42a541 881 s++;
60939fb8
NC
882
883 if (s == send)
884 return 0;
885
ae776a2c 886 /* The first digit (after optional sign): note that might
8c12dc63 887 * also point to "infinity" or "nan", or "1.#INF". */
ae776a2c
JH
888 d = s;
889
8c12dc63 890 /* next must be digit or the radix separator or beginning of infinity/nan */
60939fb8
NC
891 if (isDIGIT(*s)) {
892 /* UVs are at least 32 bits, so the first 9 decimal digits cannot
893 overflow. */
894 UV value = *s - '0';
895 /* This construction seems to be more optimiser friendly.
896 (without it gcc does the isDIGIT test and the *s - '0' separately)
897 With it gcc on arm is managing 6 instructions (6 cycles) per digit.
898 In theory the optimiser could deduce how far to unroll the loop
899 before checking for overflow. */
58bb9ec3
NC
900 if (++s < send) {
901 int 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
NC
920 if (++s < send) {
921 digit = *s - '0';
60939fb8
NC
922 if (digit >= 0 && digit <= 9) {
923 value = value * 10 + digit;
58bb9ec3
NC
924 if (++s < send) {
925 digit = *s - '0';
60939fb8
NC
926 if (digit >= 0 && digit <= 9) {
927 value = value * 10 + digit;
58bb9ec3
NC
928 if (++s < send) {
929 digit = *s - '0';
60939fb8
NC
930 if (digit >= 0 && digit <= 9) {
931 value = value * 10 + digit;
58bb9ec3 932 if (++s < send) {
60939fb8
NC
933 /* Now got 9 digits, so need to check
934 each time for overflow. */
58bb9ec3 935 digit = *s - '0';
60939fb8 936 while (digit >= 0 && digit <= 9
945b524a
JH
937 && (value < uv_max_div_10
938 || (value == uv_max_div_10
939 && digit <= uv_max_mod_10))) {
60939fb8 940 value = value * 10 + digit;
58bb9ec3
NC
941 if (++s < send)
942 digit = *s - '0';
60939fb8
NC
943 else
944 break;
945 }
946 if (digit >= 0 && digit <= 9
51bd16da 947 && (s < send)) {
60939fb8
NC
948 /* value overflowed.
949 skip the remaining digits, don't
950 worry about setting *valuep. */
951 do {
952 s++;
953 } while (s < send && isDIGIT(*s));
954 numtype |=
955 IS_NUMBER_GREATER_THAN_UV_MAX;
956 goto skip_value;
957 }
958 }
959 }
98994639 960 }
60939fb8
NC
961 }
962 }
963 }
964 }
965 }
966 }
967 }
968 }
969 }
970 }
971 }
98994639 972 }
60939fb8 973 }
98994639 974 }
60939fb8
NC
975 numtype |= IS_NUMBER_IN_UV;
976 if (valuep)
977 *valuep = value;
978
979 skip_value:
980 if (GROK_NUMERIC_RADIX(&s, send)) {
981 numtype |= IS_NUMBER_NOT_INT;
982 while (s < send && isDIGIT(*s)) /* optional digits after the radix */
983 s++;
98994639 984 }
60939fb8
NC
985 }
986 else if (GROK_NUMERIC_RADIX(&s, send)) {
987 numtype |= IS_NUMBER_NOT_INT | IS_NUMBER_IN_UV; /* valuep assigned below */
988 /* no digits before the radix means we need digits after it */
989 if (s < send && isDIGIT(*s)) {
990 do {
991 s++;
992 } while (s < send && isDIGIT(*s));
993 if (valuep) {
994 /* integer approximation is valid - it's 0. */
995 *valuep = 0;
996 }
98994639 997 }
60939fb8 998 else
ae776a2c 999 return 0;
ff4eb398 1000 }
60939fb8 1001
926f5fc6 1002 if (s > d && s < send) {
60939fb8 1003 /* we can have an optional exponent part */
305b8651 1004 if (isALPHA_FOLD_EQ(*s, 'e')) {
60939fb8
NC
1005 s++;
1006 if (s < send && (*s == '-' || *s == '+'))
1007 s++;
1008 if (s < send && isDIGIT(*s)) {
1009 do {
1010 s++;
1011 } while (s < send && isDIGIT(*s));
1012 }
3f7602fa
TC
1013 else if (flags & PERL_SCAN_TRAILING)
1014 return numtype | IS_NUMBER_TRAILING;
60939fb8 1015 else
3f7602fa
TC
1016 return 0;
1017
1018 /* The only flag we keep is sign. Blow away any "it's UV" */
1019 numtype &= IS_NUMBER_NEG;
1020 numtype |= IS_NUMBER_NOT_INT;
60939fb8
NC
1021 }
1022 }
1023 while (s < send && isSPACE(*s))
1024 s++;
1025 if (s >= send)
aa8b85de 1026 return numtype;
b59bf0b2 1027 if (memEQs(pv, len, "0 but true")) {
60939fb8
NC
1028 if (valuep)
1029 *valuep = 0;
1030 return IS_NUMBER_IN_UV;
1031 }
8c12dc63
JH
1032 /* We could be e.g. at "Inf" or "NaN", or at the "#" of "1.#INF". */
1033 if ((s + 2 < send) && strchr("inqs#", toFOLD(*s))) {
1034 /* Really detect inf/nan. Start at d, not s, since the above
1035 * code might have already consumed the "1." or "1". */
7eff3d39 1036 const int infnan = Perl_grok_infnan(aTHX_ &d, send);
8c12dc63
JH
1037 if ((infnan & IS_NUMBER_INFINITY)) {
1038 return (numtype | infnan); /* Keep sign for infinity. */
1039 }
1040 else if ((infnan & IS_NUMBER_NAN)) {
1041 return (numtype | infnan) & ~IS_NUMBER_NEG; /* Clear sign for nan. */
1042 }
1043 }
3f7602fa
TC
1044 else if (flags & PERL_SCAN_TRAILING) {
1045 return numtype | IS_NUMBER_TRAILING;
1046 }
1047
60939fb8 1048 return 0;
98994639
HS
1049}
1050
6313e544 1051/*
73e43954 1052grok_atoUV
6313e544 1053
22ff3130 1054grok_atoUV parses a C-style zero-byte terminated string, looking for
d62b8c6a 1055a decimal unsigned integer.
338aa8b0 1056
d62b8c6a
JH
1057Returns the unsigned integer, if a valid value can be parsed
1058from the beginning of the string.
f4379102 1059
d62b8c6a 1060Accepts only the decimal digits '0'..'9'.
6313e544 1061
22ff3130 1062As opposed to atoi or strtol, grok_atoUV does NOT allow optional
d62b8c6a
JH
1063leading whitespace, or negative inputs. If such features are
1064required, the calling code needs to explicitly implement those.
6313e544 1065
22ff3130
HS
1066Returns true if a valid value could be parsed. In that case, valptr
1067is set to the parsed value, and endptr (if provided) is set to point
1068to the character after the last digit.
338aa8b0 1069
22ff3130
HS
1070Returns false otherwise. This can happen if a) there is a leading zero
1071followed by another digit; b) the digits would overflow a UV; or c)
1072there are trailing non-digits AND endptr is not provided.
6313e544 1073
d62b8c6a
JH
1074Background: atoi has severe problems with illegal inputs, it cannot be
1075used for incremental parsing, and therefore should be avoided
1076atoi and strtol are also affected by locale settings, which can also be
1077seen as a bug (global state controlled by user environment).
1078
6313e544
JH
1079*/
1080
22ff3130
HS
1081bool
1082Perl_grok_atoUV(const char *pv, UV *valptr, const char** endptr)
6313e544
JH
1083{
1084 const char* s = pv;
1085 const char** eptr;
1086 const char* end2; /* Used in case endptr is NULL. */
22ff3130 1087 UV val = 0; /* The parsed value. */
6313e544 1088
22ff3130 1089 PERL_ARGS_ASSERT_GROK_ATOUV;
6313e544
JH
1090
1091 eptr = endptr ? endptr : &end2;
75feedba
JH
1092 if (isDIGIT(*s)) {
1093 /* Single-digit inputs are quite common. */
6313e544 1094 val = *s++ - '0';
75feedba 1095 if (isDIGIT(*s)) {
22ff3130
HS
1096 /* Fail on extra leading zeros. */
1097 if (val == 0)
1098 return FALSE;
75feedba
JH
1099 while (isDIGIT(*s)) {
1100 /* This could be unrolled like in grok_number(), but
1101 * the expected uses of this are not speed-needy, and
1102 * unlikely to need full 64-bitness. */
7eff3d39 1103 const U8 digit = *s++ - '0';
945b524a
JH
1104 if (val < uv_max_div_10 ||
1105 (val == uv_max_div_10 && digit <= uv_max_mod_10)) {
75feedba
JH
1106 val = val * 10 + digit;
1107 } else {
22ff3130 1108 return FALSE;
6313e544 1109 }
6313e544
JH
1110 }
1111 }
75feedba 1112 }
22ff3130
HS
1113 if (s == pv)
1114 return FALSE;
1115 if (endptr == NULL && *s)
1116 return FALSE; /* If endptr is NULL, no trailing non-digits allowed. */
6313e544 1117 *eptr = s;
22ff3130
HS
1118 *valptr = val;
1119 return TRUE;
6313e544
JH
1120}
1121
a4eca1d4 1122#ifndef USE_QUADMATH
4801ca72 1123STATIC NV
98994639
HS
1124S_mulexp10(NV value, I32 exponent)
1125{
1126 NV result = 1.0;
1127 NV power = 10.0;
1128 bool negative = 0;
1129 I32 bit;
1130
1131 if (exponent == 0)
1132 return value;
659c4b96
DM
1133 if (value == 0)
1134 return (NV)0;
87032ba1 1135
24866caa 1136 /* On OpenVMS VAX we by default use the D_FLOAT double format,
67597c89 1137 * and that format does not have *easy* capabilities [1] for
24866caa
CB
1138 * overflowing doubles 'silently' as IEEE fp does. We also need
1139 * to support G_FLOAT on both VAX and Alpha, and though the exponent
1140 * range is much larger than D_FLOAT it still doesn't do silent
1141 * overflow. Therefore we need to detect early whether we would
1142 * overflow (this is the behaviour of the native string-to-float
1143 * conversion routines, and therefore of native applications, too).
67597c89 1144 *
24866caa
CB
1145 * [1] Trying to establish a condition handler to trap floating point
1146 * exceptions is not a good idea. */
87032ba1
JH
1147
1148 /* In UNICOS and in certain Cray models (such as T90) there is no
1149 * IEEE fp, and no way at all from C to catch fp overflows gracefully.
1150 * There is something you can do if you are willing to use some
1151 * inline assembler: the instruction is called DFI-- but that will
1152 * disable *all* floating point interrupts, a little bit too large
1153 * a hammer. Therefore we need to catch potential overflows before
1154 * it's too late. */
353813d9 1155
a7157111 1156#if ((defined(VMS) && !defined(_IEEE_FP)) || defined(_UNICOS) || defined(DOUBLE_IS_VAX_FLOAT)) && defined(NV_MAX_10_EXP)
353813d9 1157 STMT_START {
c4420975 1158 const NV exp_v = log10(value);
353813d9
HS
1159 if (exponent >= NV_MAX_10_EXP || exponent + exp_v >= NV_MAX_10_EXP)
1160 return NV_MAX;
1161 if (exponent < 0) {
1162 if (-(exponent + exp_v) >= NV_MAX_10_EXP)
1163 return 0.0;
1164 while (-exponent >= NV_MAX_10_EXP) {
1165 /* combination does not overflow, but 10^(-exponent) does */
1166 value /= 10;
1167 ++exponent;
1168 }
1169 }
1170 } STMT_END;
87032ba1
JH
1171#endif
1172
353813d9
HS
1173 if (exponent < 0) {
1174 negative = 1;
1175 exponent = -exponent;
b27804d8
DM
1176#ifdef NV_MAX_10_EXP
1177 /* for something like 1234 x 10^-309, the action of calculating
1178 * the intermediate value 10^309 then returning 1234 / (10^309)
1179 * will fail, since 10^309 becomes infinity. In this case try to
1180 * refactor it as 123 / (10^308) etc.
1181 */
1182 while (value && exponent > NV_MAX_10_EXP) {
1183 exponent--;
1184 value /= 10;
1185 }
48853916
JH
1186 if (value == 0.0)
1187 return value;
b27804d8 1188#endif
353813d9 1189 }
c62e754c
JH
1190#if defined(__osf__)
1191 /* Even with cc -ieee + ieee_set_fp_control(IEEE_TRAP_ENABLE_INV)
1192 * Tru64 fp behavior on inf/nan is somewhat broken. Another way
1193 * to do this would be ieee_set_fp_control(IEEE_TRAP_ENABLE_OVF)
1194 * but that breaks another set of infnan.t tests. */
1195# define FP_OVERFLOWS_TO_ZERO
1196#endif
98994639
HS
1197 for (bit = 1; exponent; bit <<= 1) {
1198 if (exponent & bit) {
1199 exponent ^= bit;
1200 result *= power;
c62e754c
JH
1201#ifdef FP_OVERFLOWS_TO_ZERO
1202 if (result == 0)
a7157111 1203# ifdef NV_INF
c62e754c 1204 return value < 0 ? -NV_INF : NV_INF;
a7157111
JH
1205# else
1206 return value < 0 ? -FLT_MAX : FLT_MAX;
1207# endif
c62e754c 1208#endif
236f0012
CB
1209 /* Floating point exceptions are supposed to be turned off,
1210 * but if we're obviously done, don't risk another iteration.
1211 */
1212 if (exponent == 0) break;
98994639
HS
1213 }
1214 power *= power;
1215 }
1216 return negative ? value / result : value * result;
1217}
a4eca1d4 1218#endif /* #ifndef USE_QUADMATH */
98994639
HS
1219
1220NV
1221Perl_my_atof(pTHX_ const char* s)
1222{
f720c878
KW
1223 /* 's' must be NUL terminated */
1224
98994639 1225 NV x = 0.0;
9eda1ea6
KW
1226
1227 PERL_ARGS_ASSERT_MY_ATOF;
1228
a4eca1d4 1229#ifdef USE_QUADMATH
9eda1ea6 1230
a4eca1d4 1231 Perl_my_atof2(aTHX_ s, &x);
9eda1ea6
KW
1232
1233#elif ! defined(USE_LOCALE_NUMERIC)
1234
1235 Perl_atof2(s, x);
1236
1237#else
7918f24d 1238
a2287a13 1239 {
67d796ae
KW
1240 DECLARATION_FOR_LC_NUMERIC_MANIPULATION;
1241 STORE_LC_NUMERIC_SET_TO_NEEDED();
d6ded950 1242 if (PL_numeric_radix_sv && IN_LC(LC_NUMERIC)) {
e4850248
KW
1243 /* Look through the string for the first thing that looks like a
1244 * decimal point: either the value in the current locale or the
1245 * standard fallback of '.'. The one which appears earliest in the
1246 * input string is the one that we should have atof look for. Note
1247 * that we have to determine this beforehand because on some
1248 * systems, Perl_atof2 is just a wrapper around the system's atof.
1249 * */
1ae85f6c
KW
1250 const char * const standard_pos = strchr(s, '.');
1251 const char * const local_pos
1252 = strstr(s, SvPV_nolen(PL_numeric_radix_sv));
1253 const bool use_standard_radix
1254 = standard_pos && (!local_pos || standard_pos < local_pos);
78787052 1255
665873e9 1256 if (use_standard_radix) {
e4850248 1257 SET_NUMERIC_STANDARD();
665873e9
KW
1258 LOCK_LC_NUMERIC_STANDARD();
1259 }
78787052 1260
e4850248 1261 Perl_atof2(s, x);
78787052 1262
665873e9
KW
1263 if (use_standard_radix) {
1264 UNLOCK_LC_NUMERIC_STANDARD();
67d796ae 1265 SET_NUMERIC_UNDERLYING();
665873e9 1266 }
e4850248
KW
1267 }
1268 else
1269 Perl_atof2(s, x);
a2287a13
KW
1270 RESTORE_LC_NUMERIC();
1271 }
9eda1ea6 1272
98994639 1273#endif
9eda1ea6 1274
98994639
HS
1275 return x;
1276}
1277
a7157111 1278#if defined(NV_INF) || defined(NV_NAN)
3c81f0b3
DD
1279
1280#ifdef USING_MSVC6
1281# pragma warning(push)
1282# pragma warning(disable:4756;disable:4056)
1283#endif
829757a4 1284static char*
5563f457 1285S_my_atof_infnan(pTHX_ const char* s, bool negative, const char* send, NV* value)
829757a4
JH
1286{
1287 const char *p0 = negative ? s - 1 : s;
1288 const char *p = p0;
7eff3d39 1289 const int infnan = grok_infnan(&p, send);
829757a4
JH
1290 if (infnan && p != p0) {
1291 /* If we can generate inf/nan directly, let's do so. */
1292#ifdef NV_INF
1293 if ((infnan & IS_NUMBER_INFINITY)) {
3823048b 1294 *value = (infnan & IS_NUMBER_NEG) ? -NV_INF: NV_INF;
829757a4
JH
1295 return (char*)p;
1296 }
1297#endif
1298#ifdef NV_NAN
1299 if ((infnan & IS_NUMBER_NAN)) {
3823048b 1300 *value = NV_NAN;
829757a4
JH
1301 return (char*)p;
1302 }
1303#endif
1304#ifdef Perl_strtod
68611e6f 1305 /* If still here, we didn't have either NV_INF or NV_NAN,
829757a4
JH
1306 * and can try falling back to native strtod/strtold.
1307 *
1308 * The native interface might not recognize all the possible
1309 * inf/nan strings Perl recognizes. What we can try
1310 * is to try faking the input. We will try inf/-inf/nan
1311 * as the most promising/portable input. */
1312 {
1313 const char* fake = NULL;
1314 char* endp;
1315 NV nv;
a7157111 1316#ifdef NV_INF
829757a4
JH
1317 if ((infnan & IS_NUMBER_INFINITY)) {
1318 fake = ((infnan & IS_NUMBER_NEG)) ? "-inf" : "inf";
1319 }
a7157111
JH
1320#endif
1321#ifdef NV_NAN
1322 if ((infnan & IS_NUMBER_NAN)) {
829757a4
JH
1323 fake = "nan";
1324 }
a7157111 1325#endif
829757a4
JH
1326 assert(fake);
1327 nv = Perl_strtod(fake, &endp);
1328 if (fake != endp) {
a7157111 1329#ifdef NV_INF
829757a4 1330 if ((infnan & IS_NUMBER_INFINITY)) {
a7157111 1331# ifdef Perl_isinf
829757a4
JH
1332 if (Perl_isinf(nv))
1333 *value = nv;
a7157111 1334# else
829757a4
JH
1335 /* last resort, may generate SIGFPE */
1336 *value = Perl_exp((NV)1e9);
1337 if ((infnan & IS_NUMBER_NEG))
1338 *value = -*value;
a7157111 1339# endif
829757a4
JH
1340 return (char*)p; /* p, not endp */
1341 }
a7157111
JH
1342#endif
1343#ifdef NV_NAN
1344 if ((infnan & IS_NUMBER_NAN)) {
1345# ifdef Perl_isnan
829757a4
JH
1346 if (Perl_isnan(nv))
1347 *value = nv;
a7157111 1348# else
829757a4
JH
1349 /* last resort, may generate SIGFPE */
1350 *value = Perl_log((NV)-1.0);
a7157111 1351# endif
829757a4 1352 return (char*)p; /* p, not endp */
a7157111 1353#endif
829757a4
JH
1354 }
1355 }
1356 }
1357#endif /* #ifdef Perl_strtod */
1358 }
1359 return NULL;
1360}
3c81f0b3
DD
1361#ifdef USING_MSVC6
1362# pragma warning(pop)
1363#endif
829757a4 1364
a7157111
JH
1365#endif /* if defined(NV_INF) || defined(NV_NAN) */
1366
98994639
HS
1367char*
1368Perl_my_atof2(pTHX_ const char* orig, NV* value)
1369{
e1ec3a88 1370 const char* s = orig;
a4eca1d4
JH
1371 NV result[3] = {0.0, 0.0, 0.0};
1372#if defined(USE_PERL_ATOF) || defined(USE_QUADMATH)
ae776a2c 1373 const char* send = s + strlen(orig); /* one past the last */
a4eca1d4
JH
1374 bool negative = 0;
1375#endif
1376#if defined(USE_PERL_ATOF) && !defined(USE_QUADMATH)
1377 UV accumulator[2] = {0,0}; /* before/after dp */
8194bf88 1378 bool seen_digit = 0;
20f6aaab
AS
1379 I32 exp_adjust[2] = {0,0};
1380 I32 exp_acc[2] = {-1, -1};
1381 /* the current exponent adjust for the accumulators */
98994639 1382 I32 exponent = 0;
8194bf88 1383 I32 seen_dp = 0;
20f6aaab
AS
1384 I32 digit = 0;
1385 I32 old_digit = 0;
8194bf88 1386 I32 sig_digits = 0; /* noof significant digits seen so far */
a4eca1d4 1387#endif
8194bf88 1388
a4eca1d4 1389#if defined(USE_PERL_ATOF) || defined(USE_QUADMATH)
7918f24d
NC
1390 PERL_ARGS_ASSERT_MY_ATOF2;
1391
a4eca1d4
JH
1392 /* leading whitespace */
1393 while (isSPACE(*s))
1394 ++s;
1395
1396 /* sign */
1397 switch (*s) {
1398 case '-':
1399 negative = 1;
1400 /* FALLTHROUGH */
1401 case '+':
1402 ++s;
1403 }
1404#endif
1405
1406#ifdef USE_QUADMATH
1407 {
1408 char* endp;
adc55e02 1409 if ((endp = S_my_atof_infnan(aTHX_ s, negative, send, value)))
a4eca1d4
JH
1410 return endp;
1411 result[2] = strtoflt128(s, &endp);
1412 if (s != endp) {
1413 *value = negative ? -result[2] : result[2];
1414 return endp;
1415 }
1416 return NULL;
1417 }
1418#elif defined(USE_PERL_ATOF)
1419
8194bf88
DM
1420/* There is no point in processing more significant digits
1421 * than the NV can hold. Note that NV_DIG is a lower-bound value,
1422 * while we need an upper-bound value. We add 2 to account for this;
1423 * since it will have been conservative on both the first and last digit.
1424 * For example a 32-bit mantissa with an exponent of 4 would have
1425 * exact values in the set
1426 * 4
1427 * 8
1428 * ..
1429 * 17179869172
1430 * 17179869176
1431 * 17179869180
1432 *
1433 * where for the purposes of calculating NV_DIG we would have to discount
1434 * both the first and last digit, since neither can hold all values from
1435 * 0..9; but for calculating the value we must examine those two digits.
1436 */
ffa277e5
EAS
1437#ifdef MAX_SIG_DIG_PLUS
1438 /* It is not necessarily the case that adding 2 to NV_DIG gets all the
1439 possible digits in a NV, especially if NVs are not IEEE compliant
1440 (e.g., long doubles on IRIX) - Allen <allens@cpan.org> */
1441# define MAX_SIG_DIGITS (NV_DIG+MAX_SIG_DIG_PLUS)
1442#else
1443# define MAX_SIG_DIGITS (NV_DIG+2)
1444#endif
8194bf88
DM
1445
1446/* the max number we can accumulate in a UV, and still safely do 10*N+9 */
1447#define MAX_ACCUMULATE ( (UV) ((UV_MAX - 9)/10))
98994639 1448
a5dc2484 1449#if defined(NV_INF) || defined(NV_NAN)
ae776a2c 1450 {
7eff3d39 1451 char* endp;
5563f457 1452 if ((endp = S_my_atof_infnan(aTHX_ s, negative, send, value)))
7eff3d39 1453 return endp;
ae776a2c 1454 }
a5dc2484 1455#endif
2b54f59f 1456
8194bf88
DM
1457 /* we accumulate digits into an integer; when this becomes too
1458 * large, we add the total to NV and start again */
98994639 1459
8194bf88
DM
1460 while (1) {
1461 if (isDIGIT(*s)) {
1462 seen_digit = 1;
20f6aaab 1463 old_digit = digit;
8194bf88 1464 digit = *s++ - '0';
20f6aaab
AS
1465 if (seen_dp)
1466 exp_adjust[1]++;
98994639 1467
8194bf88
DM
1468 /* don't start counting until we see the first significant
1469 * digit, eg the 5 in 0.00005... */
1470 if (!sig_digits && digit == 0)
1471 continue;
1472
1473 if (++sig_digits > MAX_SIG_DIGITS) {
98994639 1474 /* limits of precision reached */
20f6aaab
AS
1475 if (digit > 5) {
1476 ++accumulator[seen_dp];
1477 } else if (digit == 5) {
1478 if (old_digit % 2) { /* round to even - Allen */
1479 ++accumulator[seen_dp];
1480 }
1481 }
1482 if (seen_dp) {
1483 exp_adjust[1]--;
1484 } else {
1485 exp_adjust[0]++;
1486 }
8194bf88 1487 /* skip remaining digits */
98994639 1488 while (isDIGIT(*s)) {
98994639 1489 ++s;
20f6aaab
AS
1490 if (! seen_dp) {
1491 exp_adjust[0]++;
1492 }
98994639
HS
1493 }
1494 /* warn of loss of precision? */
98994639 1495 }
8194bf88 1496 else {
20f6aaab 1497 if (accumulator[seen_dp] > MAX_ACCUMULATE) {
8194bf88 1498 /* add accumulator to result and start again */
20f6aaab
AS
1499 result[seen_dp] = S_mulexp10(result[seen_dp],
1500 exp_acc[seen_dp])
1501 + (NV)accumulator[seen_dp];
1502 accumulator[seen_dp] = 0;
1503 exp_acc[seen_dp] = 0;
98994639 1504 }
20f6aaab
AS
1505 accumulator[seen_dp] = accumulator[seen_dp] * 10 + digit;
1506 ++exp_acc[seen_dp];
98994639 1507 }
8194bf88 1508 }
e1ec3a88 1509 else if (!seen_dp && GROK_NUMERIC_RADIX(&s, send)) {
8194bf88 1510 seen_dp = 1;
20f6aaab 1511 if (sig_digits > MAX_SIG_DIGITS) {
9604fbf0 1512 while (isDIGIT(*s)) {
20f6aaab 1513 ++s;
9604fbf0 1514 }
20f6aaab
AS
1515 break;
1516 }
8194bf88
DM
1517 }
1518 else {
1519 break;
98994639
HS
1520 }
1521 }
1522
20f6aaab
AS
1523 result[0] = S_mulexp10(result[0], exp_acc[0]) + (NV)accumulator[0];
1524 if (seen_dp) {
1525 result[1] = S_mulexp10(result[1], exp_acc[1]) + (NV)accumulator[1];
1526 }
98994639 1527
305b8651 1528 if (seen_digit && (isALPHA_FOLD_EQ(*s, 'e'))) {
98994639
HS
1529 bool expnegative = 0;
1530
1531 ++s;
1532 switch (*s) {
1533 case '-':
1534 expnegative = 1;
924ba076 1535 /* FALLTHROUGH */
98994639
HS
1536 case '+':
1537 ++s;
1538 }
1539 while (isDIGIT(*s))
1540 exponent = exponent * 10 + (*s++ - '0');
1541 if (expnegative)
1542 exponent = -exponent;
1543 }
1544
20f6aaab
AS
1545
1546
98994639 1547 /* now apply the exponent */
20f6aaab
AS
1548
1549 if (seen_dp) {
1550 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0])
1551 + S_mulexp10(result[1],exponent-exp_adjust[1]);
1552 } else {
1553 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0]);
1554 }
98994639
HS
1555
1556 /* now apply the sign */
1557 if (negative)
20f6aaab 1558 result[2] = -result[2];
a36244b7 1559#endif /* USE_PERL_ATOF */
20f6aaab 1560 *value = result[2];
73d840c0 1561 return (char *)s;
98994639
HS
1562}
1563
5d34af89 1564/*
3d9d9213 1565=for apidoc isinfnan
5d34af89 1566
796b6530
KW
1567C<Perl_isinfnan()> is utility function that returns true if the NV
1568argument is either an infinity or a C<NaN>, false otherwise. To test
1569in more detail, use C<Perl_isinf()> and C<Perl_isnan()>.
5d34af89 1570
68611e6f
JH
1571This is also the logical inverse of Perl_isfinite().
1572
5d34af89
JH
1573=cut
1574*/
1cd88304
JH
1575bool
1576Perl_isinfnan(NV nv)
1577{
a5dc2484 1578 PERL_UNUSED_ARG(nv);
1cd88304
JH
1579#ifdef Perl_isinf
1580 if (Perl_isinf(nv))
1581 return TRUE;
1582#endif
1583#ifdef Perl_isnan
1584 if (Perl_isnan(nv))
1585 return TRUE;
1586#endif
1587 return FALSE;
1588}
1589
354b74ae
FC
1590/*
1591=for apidoc
1592
796b6530 1593Checks whether the argument would be either an infinity or C<NaN> when used
354b74ae 1594as a number, but is careful not to trigger non-numeric or uninitialized
796b6530 1595warnings. it assumes the caller has done C<SvGETMAGIC(sv)> already.
354b74ae
FC
1596
1597=cut
1598*/
1599
1600bool
1601Perl_isinfnansv(pTHX_ SV *sv)
1602{
1603 PERL_ARGS_ASSERT_ISINFNANSV;
1604 if (!SvOK(sv))
1605 return FALSE;
1606 if (SvNOKp(sv))
1607 return Perl_isinfnan(SvNVX(sv));
1608 if (SvIOKp(sv))
1609 return FALSE;
1610 {
1611 STRLEN len;
1612 const char *s = SvPV_nomg_const(sv, len);
3823048b 1613 return cBOOL(grok_infnan(&s, s+len));
354b74ae
FC
1614 }
1615}
1616
d67dac15 1617#ifndef HAS_MODFL
68611e6f
JH
1618/* C99 has truncl, pre-C99 Solaris had aintl. We can use either with
1619 * copysignl to emulate modfl, which is in some platforms missing or
1620 * broken. */
d67dac15
JH
1621# if defined(HAS_TRUNCL) && defined(HAS_COPYSIGNL)
1622long double
1623Perl_my_modfl(long double x, long double *ip)
1624{
68611e6f
JH
1625 *ip = truncl(x);
1626 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
d67dac15
JH
1627}
1628# elif defined(HAS_AINTL) && defined(HAS_COPYSIGNL)
55954f19
JH
1629long double
1630Perl_my_modfl(long double x, long double *ip)
1631{
68611e6f
JH
1632 *ip = aintl(x);
1633 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
55954f19 1634}
d67dac15 1635# endif
55954f19
JH
1636#endif
1637
7b9b7dff 1638/* Similarly, with ilogbl and scalbnl we can emulate frexpl. */
55954f19
JH
1639#if ! defined(HAS_FREXPL) && defined(HAS_ILOGBL) && defined(HAS_SCALBNL)
1640long double
1641Perl_my_frexpl(long double x, int *e) {
68611e6f
JH
1642 *e = x == 0.0L ? 0 : ilogbl(x) + 1;
1643 return (scalbnl(x, -*e));
55954f19
JH
1644}
1645#endif
66610fdd
RGS
1646
1647/*
ed140128
AD
1648=for apidoc Perl_signbit
1649
1650Return a non-zero integer if the sign bit on an NV is set, and 0 if
1651it is not.
1652
796b6530
KW
1653If F<Configure> detects this system has a C<signbit()> that will work with
1654our NVs, then we just use it via the C<#define> in F<perl.h>. Otherwise,
8b7fad81 1655fall back on this implementation. The main use of this function
796b6530 1656is catching C<-0.0>.
ed140128 1657
796b6530
KW
1658C<Configure> notes: This function is called C<'Perl_signbit'> instead of a
1659plain C<'signbit'> because it is easy to imagine a system having a C<signbit()>
ed140128 1660function or macro that doesn't happen to work with our particular choice
796b6530 1661of NVs. We shouldn't just re-C<#define> C<signbit> as C<Perl_signbit> and expect
ed140128 1662the standard system headers to be happy. Also, this is a no-context
796b6530
KW
1663function (no C<pTHX_>) because C<Perl_signbit()> is usually re-C<#defined> in
1664F<perl.h> as a simple macro call to the system's C<signbit()>.
1665Users should just always call C<Perl_signbit()>.
ed140128
AD
1666
1667=cut
1668*/
1669#if !defined(HAS_SIGNBIT)
1670int
1671Perl_signbit(NV x) {
8b7fad81 1672# ifdef Perl_fp_class_nzero
406d5545
JH
1673 return Perl_fp_class_nzero(x);
1674 /* Try finding the high byte, and assume it's highest bit
1675 * is the sign. This assumption is probably wrong somewhere. */
572cd850
JH
1676# elif defined(USE_LONG_DOUBLE) && LONG_DOUBLEKIND == LONG_DOUBLE_IS_X86_80_BIT_LITTLE_ENDIAN
1677 return (((unsigned char *)&x)[9] & 0x80);
1678# elif defined(NV_LITTLE_ENDIAN)
1679 /* Note that NVSIZE is sizeof(NV), which would make the below be
1680 * wrong if the end bytes are unused, which happens with the x86
1681 * 80-bit long doubles, which is why take care of that above. */
1682 return (((unsigned char *)&x)[NVSIZE - 1] & 0x80);
1683# elif defined(NV_BIG_ENDIAN)
1684 return (((unsigned char *)&x)[0] & 0x80);
1685# else
406d5545 1686 /* This last resort fallback is wrong for the negative zero. */
3585840c 1687 return (x < 0.0) ? 1 : 0;
572cd850 1688# endif
ed140128
AD
1689}
1690#endif
1691
1692/*
14d04a33 1693 * ex: set ts=8 sts=4 sw=4 et:
37442d52 1694 */