3 * Copyright (C) 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001,
4 * 2002, 2003, 2004, 2005, 2006, 2007, 2008 by Larry Wall and others
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.
12 * "That only makes eleven (plus one mislaid) and not fourteen,
13 * unless wizards count differently to other people." --Beorn
15 * [p.115 of _The Hobbit_: "Queer Lodgings"]
19 =head1 Numeric functions
23 This file contains all the stuff needed by perl for manipulating numeric
24 values, including such things as replacements for the OS's atof() function
29 #define PERL_IN_NUMERIC_C
36 return f < I32_MIN ? (U32) I32_MIN : (U32)(I32) f;
39 if (f < U32_MAX_P1_HALF)
42 return ((U32) f) | (1 + (U32_MAX >> 1));
47 return f > 0 ? U32_MAX : 0 /* NaN */;
54 return f < I32_MIN ? I32_MIN : (I32) f;
57 if (f < U32_MAX_P1_HALF)
60 return (I32)(((U32) f) | (1 + (U32_MAX >> 1)));
65 return f > 0 ? (I32)U32_MAX : 0 /* NaN */;
72 return f < IV_MIN ? IV_MIN : (IV) f;
75 /* For future flexibility allowing for sizeof(UV) >= sizeof(IV) */
76 if (f < UV_MAX_P1_HALF)
79 return (IV)(((UV) f) | (1 + (UV_MAX >> 1)));
84 return f > 0 ? (IV)UV_MAX : 0 /* NaN */;
91 return f < IV_MIN ? (UV) IV_MIN : (UV)(IV) f;
94 if (f < UV_MAX_P1_HALF)
97 return ((UV) f) | (1 + (UV_MAX >> 1));
102 return f > 0 ? UV_MAX : 0 /* NaN */;
108 converts a string representing a binary number to numeric form.
110 On entry C<start> and C<*len> give the string to scan, C<*flags> gives
111 conversion flags, and C<result> should be C<NULL> or a pointer to an NV.
112 The scan stops at the end of the string, or the first invalid character.
113 Unless C<PERL_SCAN_SILENT_ILLDIGIT> is set in C<*flags>, encountering an
114 invalid character will also trigger a warning.
115 On return C<*len> is set to the length of the scanned string,
116 and C<*flags> gives output flags.
118 If the value is <= C<UV_MAX> it is returned as a UV, the output flags are clear,
119 and nothing is written to C<*result>. If the value is > C<UV_MAX>, C<grok_bin>
120 returns C<UV_MAX>, sets C<PERL_SCAN_GREATER_THAN_UV_MAX> in the output flags,
121 and writes the value to C<*result> (or the value is discarded if C<result>
124 The binary number may optionally be prefixed with C<"0b"> or C<"b"> unless
125 C<PERL_SCAN_DISALLOW_PREFIX> is set in C<*flags> on entry. If
126 C<PERL_SCAN_ALLOW_UNDERSCORES> is set in C<*flags> then the binary
127 number may use C<"_"> characters to separate digits.
131 Not documented yet because experimental is C<PERL_SCAN_SILENT_NON_PORTABLE
132 which suppresses any message for non-portable numbers that are still valid
137 Perl_grok_bin(pTHX_ const char *start, STRLEN *len_p, I32 *flags, NV *result)
139 const char *s = start;
144 const UV max_div_2 = UV_MAX / 2;
145 const bool allow_underscores = cBOOL(*flags & PERL_SCAN_ALLOW_UNDERSCORES);
146 bool overflowed = FALSE;
149 PERL_ARGS_ASSERT_GROK_BIN;
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
156 if (isALPHA_FOLD_EQ(s[0], 'b')) {
160 else if (len >= 2 && s[0] == '0' && (isALPHA_FOLD_EQ(s[1], 'b'))) {
167 for (; len-- && (bit = *s); s++) {
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. */
174 if (value <= max_div_2) {
175 value = (value << 1) | (bit - '0');
178 /* Bah. We're just overflowed. */
179 /* diag_listed_as: Integer overflow in %s number */
180 Perl_ck_warner_d(aTHX_ packWARN(WARN_OVERFLOW),
181 "Integer overflow in binary number");
183 value_nv = (NV) value;
186 /* If an NV has not enough bits in its mantissa to
187 * represent a UV this summing of small low-order numbers
188 * is a waste of time (because the NV cannot preserve
189 * the low-order bits anyway): we could just remember when
190 * did we overflow and in the end just multiply value_nv by the
192 value_nv += (NV)(bit - '0');
195 if (bit == '_' && len && allow_underscores && (bit = s[1])
196 && (bit == '0' || bit == '1'))
202 if (!(*flags & PERL_SCAN_SILENT_ILLDIGIT))
203 Perl_ck_warner(aTHX_ packWARN(WARN_DIGIT),
204 "Illegal binary digit '%c' ignored", *s);
208 if ( ( overflowed && value_nv > 4294967295.0)
210 || (!overflowed && value > 0xffffffff
211 && ! (*flags & PERL_SCAN_SILENT_NON_PORTABLE))
214 Perl_ck_warner(aTHX_ packWARN(WARN_PORTABLE),
215 "Binary number > 0b11111111111111111111111111111111 non-portable");
222 *flags = PERL_SCAN_GREATER_THAN_UV_MAX;
231 converts a string representing a hex number to numeric form.
233 On entry C<start> and C<*len_p> give the string to scan, C<*flags> gives
234 conversion flags, and C<result> should be C<NULL> or a pointer to an NV.
235 The scan stops at the end of the string, or the first invalid character.
236 Unless C<PERL_SCAN_SILENT_ILLDIGIT> is set in C<*flags>, encountering an
237 invalid character will also trigger a warning.
238 On return C<*len> is set to the length of the scanned string,
239 and C<*flags> gives output flags.
241 If the value is <= C<UV_MAX> it is returned as a UV, the output flags are clear,
242 and nothing is written to C<*result>. If the value is > C<UV_MAX>, C<grok_hex>
243 returns C<UV_MAX>, sets C<PERL_SCAN_GREATER_THAN_UV_MAX> in the output flags,
244 and writes the value to C<*result> (or the value is discarded if C<result>
247 The hex number may optionally be prefixed with C<"0x"> or C<"x"> unless
248 C<PERL_SCAN_DISALLOW_PREFIX> is set in C<*flags> on entry. If
249 C<PERL_SCAN_ALLOW_UNDERSCORES> is set in C<*flags> then the hex
250 number may use C<"_"> characters to separate digits.
254 Not documented yet because experimental is C<PERL_SCAN_SILENT_NON_PORTABLE
255 which suppresses any message for non-portable numbers, but which are valid
260 Perl_grok_hex(pTHX_ const char *start, STRLEN *len_p, I32 *flags, NV *result)
262 const char *s = start;
266 const UV max_div_16 = UV_MAX / 16;
267 const bool allow_underscores = cBOOL(*flags & PERL_SCAN_ALLOW_UNDERSCORES);
268 bool overflowed = FALSE;
270 PERL_ARGS_ASSERT_GROK_HEX;
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.
277 if (isALPHA_FOLD_EQ(s[0], 'x')) {
281 else if (len >= 2 && s[0] == '0' && (isALPHA_FOLD_EQ(s[1], 'x'))) {
288 for (; len-- && *s; s++) {
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. */
295 if (value <= max_div_16) {
296 value = (value << 4) | XDIGIT_VALUE(*s);
299 /* Bah. We're just overflowed. */
300 /* diag_listed_as: Integer overflow in %s number */
301 Perl_ck_warner_d(aTHX_ packWARN(WARN_OVERFLOW),
302 "Integer overflow in hexadecimal number");
304 value_nv = (NV) value;
307 /* If an NV has not enough bits in its mantissa to
308 * represent a UV this summing of small low-order numbers
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. */
313 value_nv += (NV) XDIGIT_VALUE(*s);
316 if (*s == '_' && len && allow_underscores && s[1]
323 if (!(*flags & PERL_SCAN_SILENT_ILLDIGIT))
324 Perl_ck_warner(aTHX_ packWARN(WARN_DIGIT),
325 "Illegal hexadecimal digit '%c' ignored", *s);
329 if ( ( overflowed && value_nv > 4294967295.0)
331 || (!overflowed && value > 0xffffffff
332 && ! (*flags & PERL_SCAN_SILENT_NON_PORTABLE))
335 Perl_ck_warner(aTHX_ packWARN(WARN_PORTABLE),
336 "Hexadecimal number > 0xffffffff non-portable");
343 *flags = PERL_SCAN_GREATER_THAN_UV_MAX;
352 converts a string representing an octal number to numeric form.
354 On entry C<start> and C<*len> give the string to scan, C<*flags> gives
355 conversion flags, and C<result> should be C<NULL> or a pointer to an NV.
356 The scan stops at the end of the string, or the first invalid character.
357 Unless C<PERL_SCAN_SILENT_ILLDIGIT> is set in C<*flags>, encountering an
358 8 or 9 will also trigger a warning.
359 On return C<*len> is set to the length of the scanned string,
360 and C<*flags> gives output flags.
362 If the value is <= C<UV_MAX> it is returned as a UV, the output flags are clear,
363 and nothing is written to C<*result>. If the value is > C<UV_MAX>, C<grok_oct>
364 returns C<UV_MAX>, sets C<PERL_SCAN_GREATER_THAN_UV_MAX> in the output flags,
365 and writes the value to C<*result> (or the value is discarded if C<result>
368 If C<PERL_SCAN_ALLOW_UNDERSCORES> is set in C<*flags> then the octal
369 number may use C<"_"> characters to separate digits.
373 Not documented yet because experimental is C<PERL_SCAN_SILENT_NON_PORTABLE>
374 which suppresses any message for non-portable numbers, but which are valid
379 Perl_grok_oct(pTHX_ const char *start, STRLEN *len_p, I32 *flags, NV *result)
381 const char *s = start;
385 const UV max_div_8 = UV_MAX / 8;
386 const bool allow_underscores = cBOOL(*flags & PERL_SCAN_ALLOW_UNDERSCORES);
387 bool overflowed = FALSE;
389 PERL_ARGS_ASSERT_GROK_OCT;
391 for (; len-- && *s; s++) {
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.
398 if (value <= max_div_8) {
399 value = (value << 3) | OCTAL_VALUE(*s);
402 /* Bah. We're just overflowed. */
403 /* diag_listed_as: Integer overflow in %s number */
404 Perl_ck_warner_d(aTHX_ packWARN(WARN_OVERFLOW),
405 "Integer overflow in octal number");
407 value_nv = (NV) value;
410 /* If an NV has not enough bits in its mantissa to
411 * represent a UV this summing of small low-order numbers
412 * is a waste of time (because the NV cannot preserve
413 * the low-order bits anyway): we could just remember when
414 * did we overflow and in the end just multiply value_nv by the
415 * right amount of 8-tuples. */
416 value_nv += (NV) OCTAL_VALUE(*s);
419 if (*s == '_' && len && allow_underscores && isOCTAL(s[1])) {
424 /* Allow \octal to work the DWIM way (that is, stop scanning
425 * as soon as non-octal characters are seen, complain only if
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). */
429 if (!(*flags & PERL_SCAN_SILENT_ILLDIGIT))
430 Perl_ck_warner(aTHX_ packWARN(WARN_DIGIT),
431 "Illegal octal digit '%c' ignored", *s);
436 if ( ( overflowed && value_nv > 4294967295.0)
438 || (!overflowed && value > 0xffffffff
439 && ! (*flags & PERL_SCAN_SILENT_NON_PORTABLE))
442 Perl_ck_warner(aTHX_ packWARN(WARN_PORTABLE),
443 "Octal number > 037777777777 non-portable");
450 *flags = PERL_SCAN_GREATER_THAN_UV_MAX;
459 For backwards compatibility. Use C<grok_bin> instead.
463 For backwards compatibility. Use C<grok_hex> instead.
467 For backwards compatibility. Use C<grok_oct> instead.
473 Perl_scan_bin(pTHX_ const char *start, STRLEN len, STRLEN *retlen)
476 I32 flags = *retlen ? PERL_SCAN_ALLOW_UNDERSCORES : 0;
477 const UV ruv = grok_bin (start, &len, &flags, &rnv);
479 PERL_ARGS_ASSERT_SCAN_BIN;
482 return (flags & PERL_SCAN_GREATER_THAN_UV_MAX) ? rnv : (NV)ruv;
486 Perl_scan_oct(pTHX_ const char *start, STRLEN len, STRLEN *retlen)
489 I32 flags = *retlen ? PERL_SCAN_ALLOW_UNDERSCORES : 0;
490 const UV ruv = grok_oct (start, &len, &flags, &rnv);
492 PERL_ARGS_ASSERT_SCAN_OCT;
495 return (flags & PERL_SCAN_GREATER_THAN_UV_MAX) ? rnv : (NV)ruv;
499 Perl_scan_hex(pTHX_ const char *start, STRLEN len, STRLEN *retlen)
502 I32 flags = *retlen ? PERL_SCAN_ALLOW_UNDERSCORES : 0;
503 const UV ruv = grok_hex (start, &len, &flags, &rnv);
505 PERL_ARGS_ASSERT_SCAN_HEX;
508 return (flags & PERL_SCAN_GREATER_THAN_UV_MAX) ? rnv : (NV)ruv;
512 =for apidoc grok_numeric_radix
514 Scan and skip for a numeric decimal separator (radix).
519 Perl_grok_numeric_radix(pTHX_ const char **sp, const char *send)
521 #ifdef USE_LOCALE_NUMERIC
522 PERL_ARGS_ASSERT_GROK_NUMERIC_RADIX;
524 if (IN_LC(LC_NUMERIC)) {
525 DECLARATION_FOR_LC_NUMERIC_MANIPULATION;
526 STORE_LC_NUMERIC_SET_TO_NEEDED();
527 if (PL_numeric_radix_sv) {
529 const char * const radix = SvPV(PL_numeric_radix_sv, len);
530 if (*sp + len <= send && memEQ(*sp, radix, len)) {
532 RESTORE_LC_NUMERIC();
536 RESTORE_LC_NUMERIC();
538 /* always try "." if numeric radix didn't match because
539 * we may have data from different locales mixed */
542 PERL_ARGS_ASSERT_GROK_NUMERIC_RADIX;
544 if (*sp < send && **sp == '.') {
552 =for apidoc grok_infnan
554 Helper for C<grok_number()>, accepts various ways of spelling "infinity"
555 or "not a number", and returns one of the following flag combinations:
559 IS_NUMBER_INFINITE | IS_NUMBER_NEG
560 IS_NUMBER_NAN | IS_NUMBER_NEG
563 possibly |-ed with C<IS_NUMBER_TRAILING>.
565 If an infinity or a not-a-number is recognized, C<*sp> will point to
566 one byte past the end of the recognized string. If the recognition fails,
567 zero is returned, and C<*sp> will not move.
573 Perl_grok_infnan(pTHX_ const char** sp, const char* send)
577 #if defined(NV_INF) || defined(NV_NAN)
578 bool odh = FALSE; /* one-dot-hash: 1.#INF */
580 PERL_ARGS_ASSERT_GROK_INFNAN;
583 s++; if (s == send) return 0;
585 else if (*s == '-') {
586 flags |= IS_NUMBER_NEG; /* Yes, -NaN happens. Incorrect but happens. */
587 s++; if (s == send) return 0;
591 /* Visual C: 1.#SNAN, -1.#QNAN, 1#INF, 1.#IND (maybe also 1.#NAN)
592 * Let's keep the dot optional. */
593 s++; if (s == send) return 0;
595 s++; if (s == send) return 0;
598 s++; if (s == send) return 0;
604 if (isALPHA_FOLD_EQ(*s, 'I')) {
605 /* INF or IND (1.#IND is "indeterminate", a certain type of NAN) */
607 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
608 s++; if (s == send) return 0;
609 if (isALPHA_FOLD_EQ(*s, 'F')) {
611 if (s < send && (isALPHA_FOLD_EQ(*s, 'I'))) {
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;
620 while (*s == '0') { /* 1.#INF00 */
624 while (s < send && isSPACE(*s))
626 if (s < send && *s) {
627 flags |= IS_NUMBER_TRAILING;
629 flags |= IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT;
631 else if (isALPHA_FOLD_EQ(*s, 'D') && odh) { /* 1.#IND */
633 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
634 while (*s == '0') { /* 1.#IND00 */
638 flags |= IS_NUMBER_TRAILING;
644 /* Maybe NAN of some sort */
646 if (isALPHA_FOLD_EQ(*s, 'S') || isALPHA_FOLD_EQ(*s, 'Q')) {
648 /* XXX do something with the snan/qnan difference */
649 s++; if (s == send) return 0;
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;
657 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
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. */
671 /* C99 style "nan(123)" or Perlish equivalent "nan($uv)". */
675 return flags | IS_NUMBER_TRAILING;
678 while (t < send && *t && *t != ')') {
682 return flags | IS_NUMBER_TRAILING;
687 if (s[0] == '0' && s + 2 < t &&
688 isALPHA_FOLD_EQ(s[1], 'x') &&
691 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES;
692 nanval = grok_hex(s, &len, &flags, NULL);
693 if ((flags & PERL_SCAN_GREATER_THAN_UV_MAX)) {
696 nantype = IS_NUMBER_IN_UV;
699 } else if (s[0] == '0' && s + 2 < t &&
700 isALPHA_FOLD_EQ(s[1], 'b') &&
701 (s[2] == '0' || s[2] == '1')) {
703 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES;
704 nanval = grok_bin(s, &len, &flags, NULL);
705 if ((flags & PERL_SCAN_GREATER_THAN_UV_MAX)) {
708 nantype = IS_NUMBER_IN_UV;
714 grok_number_flags(s, t - s, &nanval,
716 PERL_SCAN_ALLOW_UNDERSCORES);
717 /* Unfortunately grok_number_flags() doesn't
718 * tell how far we got and the ')' will always
719 * be "trailing", so we need to double-check
720 * whether we had something dubious. */
721 for (u = s; u < t; u++) {
723 flags |= IS_NUMBER_TRAILING;
730 /* XXX Doesn't do octal: nan("0123").
731 * Probably not a big loss. */
733 if ((nantype & IS_NUMBER_NOT_INT) ||
734 !(nantype && IS_NUMBER_IN_UV)) {
735 /* XXX the nanval is currently unused, that is,
736 * not inserted as the NaN payload of the NV.
737 * But the above code already parses the C99
738 * nan(...) format. See below, and see also
739 * the nan() in POSIX.xs.
741 * Certain configuration combinations where
742 * NVSIZE is greater than UVSIZE mean that
743 * a single UV cannot contain all the possible
744 * NaN payload bits. There would need to be
745 * some more generic syntax than "nan($uv)".
747 * Issues to keep in mind:
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).
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. */
772 flags |= IS_NUMBER_TRAILING;
775 /* Looked like nan(...), but no close paren. */
776 flags |= IS_NUMBER_TRAILING;
779 while (s < send && isSPACE(*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
790 flags |= IS_NUMBER_TRAILING;
799 while (s < send && isSPACE(*s))
803 PERL_UNUSED_ARG(send);
804 #endif /* #if defined(NV_INF) || defined(NV_NAN) */
810 =for apidoc grok_number_flags
812 Recognise (or not) a number. The type of the number is returned
813 (0 if unrecognised), otherwise it is a bit-ORed combination of
814 C<IS_NUMBER_IN_UV>, C<IS_NUMBER_GREATER_THAN_UV_MAX>, C<IS_NUMBER_NOT_INT>,
815 C<IS_NUMBER_NEG>, C<IS_NUMBER_INFINITY>, C<IS_NUMBER_NAN> (defined in perl.h).
817 If the value of the number can fit in a UV, it is returned in C<*valuep>.
818 C<IS_NUMBER_IN_UV> will be set to indicate that C<*valuep> is valid, C<IS_NUMBER_IN_UV>
819 will never be set unless C<*valuep> is valid, but C<*valuep> may have been assigned
820 to during processing even though C<IS_NUMBER_IN_UV> is not set on return.
821 If C<valuep> is C<NULL>, C<IS_NUMBER_IN_UV> will be set for the same cases as when
822 C<valuep> is non-C<NULL>, but no actual assignment (or SEGV) will occur.
824 C<IS_NUMBER_NOT_INT> will be set with C<IS_NUMBER_IN_UV> if trailing decimals were
825 seen (in which case C<*valuep> gives the true value truncated to an integer), and
826 C<IS_NUMBER_NEG> if the number is negative (in which case C<*valuep> holds the
827 absolute value). C<IS_NUMBER_IN_UV> is not set if e notation was used or the
828 number is larger than a UV.
830 C<flags> allows only C<PERL_SCAN_TRAILING>, which allows for trailing
831 non-numeric text on an otherwise successful I<grok>, setting
832 C<IS_NUMBER_TRAILING> on the result.
834 =for apidoc grok_number
836 Identical to C<grok_number_flags()> with C<flags> set to zero.
841 Perl_grok_number(pTHX_ const char *pv, STRLEN len, UV *valuep)
843 PERL_ARGS_ASSERT_GROK_NUMBER;
845 return grok_number_flags(pv, len, valuep, 0);
848 static const UV uv_max_div_10 = UV_MAX / 10;
849 static const U8 uv_max_mod_10 = UV_MAX % 10;
852 Perl_grok_number_flags(pTHX_ const char *pv, STRLEN len, UV *valuep, U32 flags)
855 const char * const send = pv + len;
859 PERL_ARGS_ASSERT_GROK_NUMBER_FLAGS;
861 while (s < send && isSPACE(*s))
865 } else if (*s == '-') {
867 numtype = IS_NUMBER_NEG;
875 /* The first digit (after optional sign): note that might
876 * also point to "infinity" or "nan", or "1.#INF". */
879 /* next must be digit or the radix separator or beginning of infinity/nan */
881 /* UVs are at least 32 bits, so the first 9 decimal digits cannot
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. */
890 int digit = *s - '0';
891 if (digit >= 0 && digit <= 9) {
892 value = value * 10 + digit;
895 if (digit >= 0 && digit <= 9) {
896 value = value * 10 + digit;
899 if (digit >= 0 && digit <= 9) {
900 value = value * 10 + digit;
903 if (digit >= 0 && digit <= 9) {
904 value = value * 10 + digit;
907 if (digit >= 0 && digit <= 9) {
908 value = value * 10 + digit;
911 if (digit >= 0 && digit <= 9) {
912 value = value * 10 + digit;
915 if (digit >= 0 && digit <= 9) {
916 value = value * 10 + digit;
919 if (digit >= 0 && digit <= 9) {
920 value = value * 10 + digit;
922 /* Now got 9 digits, so need to check
923 each time for overflow. */
925 while (digit >= 0 && digit <= 9
926 && (value < uv_max_div_10
927 || (value == uv_max_div_10
928 && digit <= uv_max_mod_10))) {
929 value = value * 10 + digit;
935 if (digit >= 0 && digit <= 9
938 skip the remaining digits, don't
939 worry about setting *valuep. */
942 } while (s < send && isDIGIT(*s));
944 IS_NUMBER_GREATER_THAN_UV_MAX;
964 numtype |= IS_NUMBER_IN_UV;
969 if (GROK_NUMERIC_RADIX(&s, send)) {
970 numtype |= IS_NUMBER_NOT_INT;
971 while (s < send && isDIGIT(*s)) /* optional digits after the radix */
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)) {
981 } while (s < send && isDIGIT(*s));
983 /* integer approximation is valid - it's 0. */
991 if (s > d && s < send) {
992 /* we can have an optional exponent part */
993 if (isALPHA_FOLD_EQ(*s, 'e')) {
995 if (s < send && (*s == '-' || *s == '+'))
997 if (s < send && isDIGIT(*s)) {
1000 } while (s < send && isDIGIT(*s));
1002 else if (flags & PERL_SCAN_TRAILING)
1003 return numtype | IS_NUMBER_TRAILING;
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;
1012 while (s < send && isSPACE(*s))
1016 if (len == 10 && _memEQs(pv, "0 but true")) {
1019 return IS_NUMBER_IN_UV;
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". */
1025 const int infnan = Perl_grok_infnan(aTHX_ &d, send);
1026 if ((infnan & IS_NUMBER_INFINITY)) {
1027 return (numtype | infnan); /* Keep sign for infinity. */
1029 else if ((infnan & IS_NUMBER_NAN)) {
1030 return (numtype | infnan) & ~IS_NUMBER_NEG; /* Clear sign for nan. */
1033 else if (flags & PERL_SCAN_TRAILING) {
1034 return numtype | IS_NUMBER_TRAILING;
1043 grok_atoUV parses a C-style zero-byte terminated string, looking for
1044 a decimal unsigned integer.
1046 Returns the unsigned integer, if a valid value can be parsed
1047 from the beginning of the string.
1049 Accepts only the decimal digits '0'..'9'.
1051 As opposed to atoi or strtol, grok_atoUV does NOT allow optional
1052 leading whitespace, or negative inputs. If such features are
1053 required, the calling code needs to explicitly implement those.
1055 Returns true if a valid value could be parsed. In that case, valptr
1056 is set to the parsed value, and endptr (if provided) is set to point
1057 to the character after the last digit.
1059 Returns false otherwise. This can happen if a) there is a leading zero
1060 followed by another digit; b) the digits would overflow a UV; or c)
1061 there are trailing non-digits AND endptr is not provided.
1063 Background: atoi has severe problems with illegal inputs, it cannot be
1064 used for incremental parsing, and therefore should be avoided
1065 atoi and strtol are also affected by locale settings, which can also be
1066 seen as a bug (global state controlled by user environment).
1071 Perl_grok_atoUV(const char *pv, UV *valptr, const char** endptr)
1075 const char* end2; /* Used in case endptr is NULL. */
1076 UV val = 0; /* The parsed value. */
1078 PERL_ARGS_ASSERT_GROK_ATOUV;
1080 eptr = endptr ? endptr : &end2;
1082 /* Single-digit inputs are quite common. */
1085 /* Fail on extra leading zeros. */
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 const U8 digit = *s++ - '0';
1093 if (val < uv_max_div_10 ||
1094 (val == uv_max_div_10 && digit <= uv_max_mod_10)) {
1095 val = val * 10 + digit;
1104 if (endptr == NULL && *s)
1105 return FALSE; /* If endptr is NULL, no trailing non-digits allowed. */
1111 #ifndef USE_QUADMATH
1113 S_mulexp10(NV value, I32 exponent)
1125 /* On OpenVMS VAX we by default use the D_FLOAT double format,
1126 * and that format does not have *easy* capabilities [1] for
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).
1134 * [1] Trying to establish a condition handler to trap floating point
1135 * exceptions is not a good idea. */
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
1145 #if ((defined(VMS) && !defined(_IEEE_FP)) || defined(_UNICOS) || defined(DOUBLE_IS_VAX_FLOAT)) && defined(NV_MAX_10_EXP)
1147 const NV exp_v = log10(value);
1148 if (exponent >= NV_MAX_10_EXP || exponent + exp_v >= NV_MAX_10_EXP)
1151 if (-(exponent + exp_v) >= NV_MAX_10_EXP)
1153 while (-exponent >= NV_MAX_10_EXP) {
1154 /* combination does not overflow, but 10^(-exponent) does */
1164 exponent = -exponent;
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.
1171 while (value && exponent > NV_MAX_10_EXP) {
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
1186 for (bit = 1; exponent; bit <<= 1) {
1187 if (exponent & bit) {
1190 #ifdef FP_OVERFLOWS_TO_ZERO
1193 return value < 0 ? -NV_INF : NV_INF;
1195 return value < 0 ? -FLT_MAX : FLT_MAX;
1198 /* Floating point exceptions are supposed to be turned off,
1199 * but if we're obviously done, don't risk another iteration.
1201 if (exponent == 0) break;
1205 return negative ? value / result : value * result;
1207 #endif /* #ifndef USE_QUADMATH */
1210 Perl_my_atof(pTHX_ const char* s)
1214 Perl_my_atof2(aTHX_ s, &x);
1217 # ifdef USE_LOCALE_NUMERIC
1218 PERL_ARGS_ASSERT_MY_ATOF;
1221 DECLARATION_FOR_LC_NUMERIC_MANIPULATION;
1222 STORE_LC_NUMERIC_SET_TO_NEEDED();
1223 if (PL_numeric_radix_sv && IN_LC(LC_NUMERIC)) {
1224 /* Look through the string for the first thing that looks like a
1225 * decimal point: either the value in the current locale or the
1226 * standard fallback of '.'. The one which appears earliest in the
1227 * input string is the one that we should have atof look for. Note
1228 * that we have to determine this beforehand because on some
1229 * systems, Perl_atof2 is just a wrapper around the system's atof.
1231 const char * const standard = strchr(s, '.');
1232 const char * const local = strstr(s, SvPV_nolen(PL_numeric_radix_sv));
1233 const bool use_standard_radix = standard && (!local || standard < local);
1235 if (use_standard_radix)
1236 SET_NUMERIC_STANDARD();
1240 if (use_standard_radix)
1241 SET_NUMERIC_UNDERLYING();
1245 RESTORE_LC_NUMERIC();
1254 #if defined(NV_INF) || defined(NV_NAN)
1257 # pragma warning(push)
1258 # pragma warning(disable:4756;disable:4056)
1261 S_my_atof_infnan(pTHX_ const char* s, bool negative, const char* send, NV* value)
1263 const char *p0 = negative ? s - 1 : s;
1265 const int infnan = grok_infnan(&p, send);
1266 if (infnan && p != p0) {
1267 /* If we can generate inf/nan directly, let's do so. */
1269 if ((infnan & IS_NUMBER_INFINITY)) {
1270 *value = (infnan & IS_NUMBER_NEG) ? -NV_INF: NV_INF;
1275 if ((infnan & IS_NUMBER_NAN)) {
1281 /* If still here, we didn't have either NV_INF or NV_NAN,
1282 * and can try falling back to native strtod/strtold.
1284 * The native interface might not recognize all the possible
1285 * inf/nan strings Perl recognizes. What we can try
1286 * is to try faking the input. We will try inf/-inf/nan
1287 * as the most promising/portable input. */
1289 const char* fake = NULL;
1293 if ((infnan & IS_NUMBER_INFINITY)) {
1294 fake = ((infnan & IS_NUMBER_NEG)) ? "-inf" : "inf";
1298 if ((infnan & IS_NUMBER_NAN)) {
1303 nv = Perl_strtod(fake, &endp);
1306 if ((infnan & IS_NUMBER_INFINITY)) {
1311 /* last resort, may generate SIGFPE */
1312 *value = Perl_exp((NV)1e9);
1313 if ((infnan & IS_NUMBER_NEG))
1316 return (char*)p; /* p, not endp */
1320 if ((infnan & IS_NUMBER_NAN)) {
1325 /* last resort, may generate SIGFPE */
1326 *value = Perl_log((NV)-1.0);
1328 return (char*)p; /* p, not endp */
1333 #endif /* #ifdef Perl_strtod */
1338 # pragma warning(pop)
1341 #endif /* if defined(NV_INF) || defined(NV_NAN) */
1344 Perl_my_atof2(pTHX_ const char* orig, NV* value)
1346 const char* s = orig;
1347 NV result[3] = {0.0, 0.0, 0.0};
1348 #if defined(USE_PERL_ATOF) || defined(USE_QUADMATH)
1349 const char* send = s + strlen(orig); /* one past the last */
1352 #if defined(USE_PERL_ATOF) && !defined(USE_QUADMATH)
1353 UV accumulator[2] = {0,0}; /* before/after dp */
1354 bool seen_digit = 0;
1355 I32 exp_adjust[2] = {0,0};
1356 I32 exp_acc[2] = {-1, -1};
1357 /* the current exponent adjust for the accumulators */
1362 I32 sig_digits = 0; /* noof significant digits seen so far */
1365 #if defined(USE_PERL_ATOF) || defined(USE_QUADMATH)
1366 PERL_ARGS_ASSERT_MY_ATOF2;
1368 /* leading whitespace */
1385 if ((endp = S_my_atof_infnan(aTHX_ s, negative, send, value)))
1387 result[2] = strtoflt128(s, &endp);
1389 *value = negative ? -result[2] : result[2];
1394 #elif defined(USE_PERL_ATOF)
1396 /* There is no point in processing more significant digits
1397 * than the NV can hold. Note that NV_DIG is a lower-bound value,
1398 * while we need an upper-bound value. We add 2 to account for this;
1399 * since it will have been conservative on both the first and last digit.
1400 * For example a 32-bit mantissa with an exponent of 4 would have
1401 * exact values in the set
1409 * where for the purposes of calculating NV_DIG we would have to discount
1410 * both the first and last digit, since neither can hold all values from
1411 * 0..9; but for calculating the value we must examine those two digits.
1413 #ifdef MAX_SIG_DIG_PLUS
1414 /* It is not necessarily the case that adding 2 to NV_DIG gets all the
1415 possible digits in a NV, especially if NVs are not IEEE compliant
1416 (e.g., long doubles on IRIX) - Allen <allens@cpan.org> */
1417 # define MAX_SIG_DIGITS (NV_DIG+MAX_SIG_DIG_PLUS)
1419 # define MAX_SIG_DIGITS (NV_DIG+2)
1422 /* the max number we can accumulate in a UV, and still safely do 10*N+9 */
1423 #define MAX_ACCUMULATE ( (UV) ((UV_MAX - 9)/10))
1425 #if defined(NV_INF) || defined(NV_NAN)
1428 if ((endp = S_my_atof_infnan(aTHX_ s, negative, send, value)))
1433 /* we accumulate digits into an integer; when this becomes too
1434 * large, we add the total to NV and start again */
1444 /* don't start counting until we see the first significant
1445 * digit, eg the 5 in 0.00005... */
1446 if (!sig_digits && digit == 0)
1449 if (++sig_digits > MAX_SIG_DIGITS) {
1450 /* limits of precision reached */
1452 ++accumulator[seen_dp];
1453 } else if (digit == 5) {
1454 if (old_digit % 2) { /* round to even - Allen */
1455 ++accumulator[seen_dp];
1463 /* skip remaining digits */
1464 while (isDIGIT(*s)) {
1470 /* warn of loss of precision? */
1473 if (accumulator[seen_dp] > MAX_ACCUMULATE) {
1474 /* add accumulator to result and start again */
1475 result[seen_dp] = S_mulexp10(result[seen_dp],
1477 + (NV)accumulator[seen_dp];
1478 accumulator[seen_dp] = 0;
1479 exp_acc[seen_dp] = 0;
1481 accumulator[seen_dp] = accumulator[seen_dp] * 10 + digit;
1485 else if (!seen_dp && GROK_NUMERIC_RADIX(&s, send)) {
1487 if (sig_digits > MAX_SIG_DIGITS) {
1490 } while (isDIGIT(*s));
1499 result[0] = S_mulexp10(result[0], exp_acc[0]) + (NV)accumulator[0];
1501 result[1] = S_mulexp10(result[1], exp_acc[1]) + (NV)accumulator[1];
1504 if (seen_digit && (isALPHA_FOLD_EQ(*s, 'e'))) {
1505 bool expnegative = 0;
1516 exponent = exponent * 10 + (*s++ - '0');
1518 exponent = -exponent;
1523 /* now apply the exponent */
1526 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0])
1527 + S_mulexp10(result[1],exponent-exp_adjust[1]);
1529 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0]);
1532 /* now apply the sign */
1534 result[2] = -result[2];
1535 #endif /* USE_PERL_ATOF */
1541 =for apidoc isinfnan
1543 C<Perl_isinfnan()> is utility function that returns true if the NV
1544 argument is either an infinity or a C<NaN>, false otherwise. To test
1545 in more detail, use C<Perl_isinf()> and C<Perl_isnan()>.
1547 This is also the logical inverse of Perl_isfinite().
1552 Perl_isinfnan(NV nv)
1554 PERL_UNUSED_ARG(nv);
1569 Checks whether the argument would be either an infinity or C<NaN> when used
1570 as a number, but is careful not to trigger non-numeric or uninitialized
1571 warnings. it assumes the caller has done C<SvGETMAGIC(sv)> already.
1577 Perl_isinfnansv(pTHX_ SV *sv)
1579 PERL_ARGS_ASSERT_ISINFNANSV;
1583 return Perl_isinfnan(SvNVX(sv));
1588 const char *s = SvPV_nomg_const(sv, len);
1589 return cBOOL(grok_infnan(&s, s+len));
1594 /* C99 has truncl, pre-C99 Solaris had aintl. We can use either with
1595 * copysignl to emulate modfl, which is in some platforms missing or
1597 # if defined(HAS_TRUNCL) && defined(HAS_COPYSIGNL)
1599 Perl_my_modfl(long double x, long double *ip)
1602 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
1604 # elif defined(HAS_AINTL) && defined(HAS_COPYSIGNL)
1606 Perl_my_modfl(long double x, long double *ip)
1609 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
1614 /* Similarly, with ilogbl and scalbnl we can emulate frexpl. */
1615 #if ! defined(HAS_FREXPL) && defined(HAS_ILOGBL) && defined(HAS_SCALBNL)
1617 Perl_my_frexpl(long double x, int *e) {
1618 *e = x == 0.0L ? 0 : ilogbl(x) + 1;
1619 return (scalbnl(x, -*e));
1624 =for apidoc Perl_signbit
1626 Return a non-zero integer if the sign bit on an NV is set, and 0 if
1629 If F<Configure> detects this system has a C<signbit()> that will work with
1630 our NVs, then we just use it via the C<#define> in F<perl.h>. Otherwise,
1631 fall back on this implementation. The main use of this function
1632 is catching C<-0.0>.
1634 C<Configure> notes: This function is called C<'Perl_signbit'> instead of a
1635 plain C<'signbit'> because it is easy to imagine a system having a C<signbit()>
1636 function or macro that doesn't happen to work with our particular choice
1637 of NVs. We shouldn't just re-C<#define> C<signbit> as C<Perl_signbit> and expect
1638 the standard system headers to be happy. Also, this is a no-context
1639 function (no C<pTHX_>) because C<Perl_signbit()> is usually re-C<#defined> in
1640 F<perl.h> as a simple macro call to the system's C<signbit()>.
1641 Users should just always call C<Perl_signbit()>.
1645 #if !defined(HAS_SIGNBIT)
1647 Perl_signbit(NV x) {
1648 # ifdef Perl_fp_class_nzero
1649 return Perl_fp_class_nzero(x);
1650 /* Try finding the high byte, and assume it's highest bit
1651 * is the sign. This assumption is probably wrong somewhere. */
1652 # elif defined(USE_LONG_DOUBLE) && LONG_DOUBLEKIND == LONG_DOUBLE_IS_X86_80_BIT_LITTLE_ENDIAN
1653 return (((unsigned char *)&x)[9] & 0x80);
1654 # elif defined(NV_LITTLE_ENDIAN)
1655 /* Note that NVSIZE is sizeof(NV), which would make the below be
1656 * wrong if the end bytes are unused, which happens with the x86
1657 * 80-bit long doubles, which is why take care of that above. */
1658 return (((unsigned char *)&x)[NVSIZE - 1] & 0x80);
1659 # elif defined(NV_BIG_ENDIAN)
1660 return (((unsigned char *)&x)[0] & 0x80);
1662 /* This last resort fallback is wrong for the negative zero. */
1663 return (x < 0.0) ? 1 : 0;
1669 * ex: set ts=8 sts=4 sw=4 et: