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 I<start> and I<*len> give the string to scan, I<*flags> gives
111 conversion flags, and I<result> should be 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 I<*flags>, encountering an
114 invalid character will also trigger a warning.
115 On return I<*len> is set to the length of the scanned string,
116 and I<*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 I<*result>. If the value is > UV_MAX C<grok_bin>
120 returns UV_MAX, sets C<PERL_SCAN_GREATER_THAN_UV_MAX> in the output flags,
121 and writes the value to I<*result> (or the value is discarded if I<result>
124 The binary number may optionally be prefixed with "0b" or "b" unless
125 C<PERL_SCAN_DISALLOW_PREFIX> is set in I<*flags> on entry. If
126 C<PERL_SCAN_ALLOW_UNDERSCORES> is set in I<*flags> then the binary
127 number may use '_' 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 I<start> and I<*len_p> give the string to scan, I<*flags> gives
234 conversion flags, and I<result> should be 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 I<*flags>, encountering an
237 invalid character will also trigger a warning.
238 On return I<*len> is set to the length of the scanned string,
239 and I<*flags> gives output flags.
241 If the value is <= UV_MAX it is returned as a UV, the output flags are clear,
242 and nothing is written to I<*result>. If the value is > UV_MAX C<grok_hex>
243 returns UV_MAX, sets C<PERL_SCAN_GREATER_THAN_UV_MAX> in the output flags,
244 and writes the value to I<*result> (or the value is discarded if I<result>
247 The hex number may optionally be prefixed with "0x" or "x" unless
248 C<PERL_SCAN_DISALLOW_PREFIX> is set in I<*flags> on entry. If
249 C<PERL_SCAN_ALLOW_UNDERSCORES> is set in I<*flags> then the hex
250 number may use '_' 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 I<start> and I<*len> give the string to scan, I<*flags> gives
355 conversion flags, and I<result> should be 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 I<*flags>, encountering an
358 8 or 9 will also trigger a warning.
359 On return I<*len> is set to the length of the scanned string,
360 and I<*flags> gives output flags.
362 If the value is <= UV_MAX it is returned as a UV, the output flags are clear,
363 and nothing is written to I<*result>. If the value is > UV_MAX C<grok_oct>
364 returns UV_MAX, sets C<PERL_SCAN_GREATER_THAN_UV_MAX> in the output flags,
365 and writes the value to I<*result> (or the value is discarded if I<result>
368 If C<PERL_SCAN_ALLOW_UNDERSCORES> is set in I<*flags> then the octal
369 number may use '_' 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 DECLARE_STORE_LC_NUMERIC_SET_TO_NEEDED();
526 if (PL_numeric_radix_sv) {
528 const char * const radix = SvPV(PL_numeric_radix_sv, len);
529 if (*sp + len <= send && memEQ(*sp, radix, len)) {
531 RESTORE_LC_NUMERIC();
535 RESTORE_LC_NUMERIC();
537 /* always try "." if numeric radix didn't match because
538 * we may have data from different locales mixed */
541 PERL_ARGS_ASSERT_GROK_NUMERIC_RADIX;
543 if (*sp < send && **sp == '.') {
551 =for apidoc grok_infnan
553 Helper for grok_number(), accepts various ways of spelling "infinity"
554 or "not a number", and returns one of the following flag combinations:
558 IS_NUMBER_INFINITE | IS_NUMBER_NEG
559 IS_NUMBER_NAN | IS_NUMBER_NEG
562 possibly |-ed with IS_NUMBER_TRAILING.
564 If an infinity or a not-a-number is recognized, the *sp will point to
565 one byte past the end of the recognized string. If the recognition fails,
566 zero is returned, and the *sp will not move.
572 Perl_grok_infnan(pTHX_ const char** sp, const char* send)
576 bool odh = FALSE; /* one-dot-hash: 1.#INF */
578 PERL_ARGS_ASSERT_GROK_INFNAN;
581 s++; if (s == send) return 0;
583 else if (*s == '-') {
584 flags |= IS_NUMBER_NEG; /* Yes, -NaN happens. Incorrect but happens. */
585 s++; if (s == send) return 0;
589 /* Visual C: 1.#SNAN, -1.#QNAN, 1#INF, 1.#IND (maybe also 1.#NAN)
590 * Let's keep the dot optional. */
591 s++; if (s == send) return 0;
593 s++; if (s == send) return 0;
596 s++; if (s == send) return 0;
602 if (isALPHA_FOLD_EQ(*s, 'I')) {
603 /* INF or IND (1.#IND is "indeterminate", a certain type of NAN) */
605 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
606 s++; if (s == send) return 0;
607 if (isALPHA_FOLD_EQ(*s, 'F')) {
609 while (*s == '0') { /* 1.#INF00 */
612 while (s < send && isSPACE(*s))
614 if (s < send && *s) {
615 flags |= IS_NUMBER_TRAILING;
617 flags |= IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT;
619 else if (isALPHA_FOLD_EQ(*s, 'D') && odh) { /* 1.#IND */
621 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
622 while (*s == '0') { /* 1.#IND00 */
626 flags |= IS_NUMBER_TRAILING;
632 /* Maybe NAN of some sort */
634 if (isALPHA_FOLD_EQ(*s, 'S') || isALPHA_FOLD_EQ(*s, 'Q')) {
636 /* XXX do something with the snan/qnan difference */
637 s++; if (s == send) return 0;
640 if (isALPHA_FOLD_EQ(*s, 'N')) {
641 s++; if (s == send || isALPHA_FOLD_NE(*s, 'A')) return 0;
642 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
645 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
647 /* NaN can be followed by various stuff (NaNQ, NaNS), but
648 * there are also multiple different NaN values, and some
649 * implementations output the "payload" values,
650 * e.g. NaN123, NAN(abc), while some legacy implementations
651 * have weird stuff like NaN%. */
652 if (isALPHA_FOLD_EQ(*s, 'q') ||
653 isALPHA_FOLD_EQ(*s, 's')) {
654 /* "nanq" or "nans" are ok, though generating
655 * these portably is tricky. */
659 /* C99 style "nan(123)" or Perlish equivalent "nan($uv)". */
663 return flags | IS_NUMBER_TRAILING;
666 while (t < send && *t && *t != ')') {
670 return flags | IS_NUMBER_TRAILING;
675 if (s[0] == '0' && s + 2 < t &&
676 isALPHA_FOLD_EQ(s[1], 'x') &&
679 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES;
680 nanval = grok_hex(s, &len, &flags, NULL);
681 if ((flags & PERL_SCAN_GREATER_THAN_UV_MAX)) {
684 nantype = IS_NUMBER_IN_UV;
687 } else if (s[0] == '0' && s + 2 < t &&
688 isALPHA_FOLD_EQ(s[1], 'b') &&
689 (s[2] == '0' || s[2] == '1')) {
691 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES;
692 nanval = grok_bin(s, &len, &flags, NULL);
693 if ((flags & PERL_SCAN_GREATER_THAN_UV_MAX)) {
696 nantype = IS_NUMBER_IN_UV;
702 grok_number_flags(s, t - s, &nanval,
704 PERL_SCAN_ALLOW_UNDERSCORES);
705 /* Unfortunately grok_number_flags() doesn't
706 * tell how far we got and the ')' will always
707 * be "trailing", so we need to double-check
708 * whether we had something dubious. */
709 for (u = s; u < t; u++) {
711 flags |= IS_NUMBER_TRAILING;
718 /* "What about octal?" Really? */
720 if ((nantype & IS_NUMBER_NOT_INT) ||
721 !(nantype && IS_NUMBER_IN_UV)) {
722 /* Certain configuration combinations where
723 * NVSIZE is greater than UVSIZE mean that a
724 * single UV cannot contain all the possible
725 * NaN payload bits. There would need to be
726 * some more generic syntax than "nan($uv)".
727 * Issues to keep in mind:
728 * (1) In most common cases there would
729 * not be an integral number of bytes that
730 * could be set, only a certain number of bits.
731 * For example for NVSIZE==UVSIZE there can be
732 * up to 52 bits in the payload, but one bit is
733 * commonly reserved for the signal/quiet bit,
735 * (2) Endianness of the payload bits. If the
736 * payload is specified as an UV, the low-order
737 * bits of the UV are naturally little-endianed
738 * (rightmost) bits of the payload. */
742 flags |= IS_NUMBER_TRAILING;
745 /* Looked like nan(...), but no close paren. */
746 flags |= IS_NUMBER_TRAILING;
749 while (s < send && isSPACE(*s))
751 if (s < send && *s) {
752 /* Note that we here implicitly accept (parse as
753 * "nan", but with warnings) also any other weird
754 * trailing stuff for "nan". In the above we just
755 * check that if we got the C99-style "nan(...)",
756 * the "..." looks sane.
757 * If in future we accept more ways of specifying
758 * the nan payload, the accepting would happen around
760 flags |= IS_NUMBER_TRAILING;
769 while (s < send && isSPACE(*s))
777 =for apidoc grok_number_flags
779 Recognise (or not) a number. The type of the number is returned
780 (0 if unrecognised), otherwise it is a bit-ORed combination of
781 IS_NUMBER_IN_UV, IS_NUMBER_GREATER_THAN_UV_MAX, IS_NUMBER_NOT_INT,
782 IS_NUMBER_NEG, IS_NUMBER_INFINITY, IS_NUMBER_NAN (defined in perl.h).
784 If the value of the number can fit in a UV, it is returned in the *valuep
785 IS_NUMBER_IN_UV will be set to indicate that *valuep is valid, IS_NUMBER_IN_UV
786 will never be set unless *valuep is valid, but *valuep may have been assigned
787 to during processing even though IS_NUMBER_IN_UV is not set on return.
788 If valuep is NULL, IS_NUMBER_IN_UV will be set for the same cases as when
789 valuep is non-NULL, but no actual assignment (or SEGV) will occur.
791 IS_NUMBER_NOT_INT will be set with IS_NUMBER_IN_UV if trailing decimals were
792 seen (in which case *valuep gives the true value truncated to an integer), and
793 IS_NUMBER_NEG if the number is negative (in which case *valuep holds the
794 absolute value). IS_NUMBER_IN_UV is not set if e notation was used or the
795 number is larger than a UV.
797 C<flags> allows only C<PERL_SCAN_TRAILING>, which allows for trailing
798 non-numeric text on an otherwise successful I<grok>, setting
799 C<IS_NUMBER_TRAILING> on the result.
801 =for apidoc grok_number
803 Identical to grok_number_flags() with flags set to zero.
808 Perl_grok_number(pTHX_ const char *pv, STRLEN len, UV *valuep)
810 PERL_ARGS_ASSERT_GROK_NUMBER;
812 return grok_number_flags(pv, len, valuep, 0);
815 static const UV uv_max_div_10 = UV_MAX / 10;
816 static const U8 uv_max_mod_10 = UV_MAX % 10;
819 Perl_grok_number_flags(pTHX_ const char *pv, STRLEN len, UV *valuep, U32 flags)
822 const char * const send = pv + len;
826 PERL_ARGS_ASSERT_GROK_NUMBER_FLAGS;
828 while (s < send && isSPACE(*s))
832 } else if (*s == '-') {
834 numtype = IS_NUMBER_NEG;
842 /* The first digit (after optional sign): note that might
843 * also point to "infinity" or "nan", or "1.#INF". */
846 /* next must be digit or the radix separator or beginning of infinity/nan */
848 /* UVs are at least 32 bits, so the first 9 decimal digits cannot
851 /* This construction seems to be more optimiser friendly.
852 (without it gcc does the isDIGIT test and the *s - '0' separately)
853 With it gcc on arm is managing 6 instructions (6 cycles) per digit.
854 In theory the optimiser could deduce how far to unroll the loop
855 before checking for overflow. */
857 int digit = *s - '0';
858 if (digit >= 0 && digit <= 9) {
859 value = value * 10 + digit;
862 if (digit >= 0 && digit <= 9) {
863 value = value * 10 + digit;
866 if (digit >= 0 && digit <= 9) {
867 value = value * 10 + digit;
870 if (digit >= 0 && digit <= 9) {
871 value = value * 10 + digit;
874 if (digit >= 0 && digit <= 9) {
875 value = value * 10 + digit;
878 if (digit >= 0 && digit <= 9) {
879 value = value * 10 + digit;
882 if (digit >= 0 && digit <= 9) {
883 value = value * 10 + digit;
886 if (digit >= 0 && digit <= 9) {
887 value = value * 10 + digit;
889 /* Now got 9 digits, so need to check
890 each time for overflow. */
892 while (digit >= 0 && digit <= 9
893 && (value < uv_max_div_10
894 || (value == uv_max_div_10
895 && digit <= uv_max_mod_10))) {
896 value = value * 10 + digit;
902 if (digit >= 0 && digit <= 9
905 skip the remaining digits, don't
906 worry about setting *valuep. */
909 } while (s < send && isDIGIT(*s));
911 IS_NUMBER_GREATER_THAN_UV_MAX;
931 numtype |= IS_NUMBER_IN_UV;
936 if (GROK_NUMERIC_RADIX(&s, send)) {
937 numtype |= IS_NUMBER_NOT_INT;
938 while (s < send && isDIGIT(*s)) /* optional digits after the radix */
942 else if (GROK_NUMERIC_RADIX(&s, send)) {
943 numtype |= IS_NUMBER_NOT_INT | IS_NUMBER_IN_UV; /* valuep assigned below */
944 /* no digits before the radix means we need digits after it */
945 if (s < send && isDIGIT(*s)) {
948 } while (s < send && isDIGIT(*s));
950 /* integer approximation is valid - it's 0. */
958 if (s > d && s < send) {
959 /* we can have an optional exponent part */
960 if (isALPHA_FOLD_EQ(*s, 'e')) {
962 if (s < send && (*s == '-' || *s == '+'))
964 if (s < send && isDIGIT(*s)) {
967 } while (s < send && isDIGIT(*s));
969 else if (flags & PERL_SCAN_TRAILING)
970 return numtype | IS_NUMBER_TRAILING;
974 /* The only flag we keep is sign. Blow away any "it's UV" */
975 numtype &= IS_NUMBER_NEG;
976 numtype |= IS_NUMBER_NOT_INT;
979 while (s < send && isSPACE(*s))
983 if (len == 10 && memEQ(pv, "0 but true", 10)) {
986 return IS_NUMBER_IN_UV;
988 /* We could be e.g. at "Inf" or "NaN", or at the "#" of "1.#INF". */
989 if ((s + 2 < send) && strchr("inqs#", toFOLD(*s))) {
990 /* Really detect inf/nan. Start at d, not s, since the above
991 * code might have already consumed the "1." or "1". */
992 int infnan = Perl_grok_infnan(aTHX_ &d, send);
993 if ((infnan & IS_NUMBER_INFINITY)) {
994 return (numtype | infnan); /* Keep sign for infinity. */
996 else if ((infnan & IS_NUMBER_NAN)) {
997 return (numtype | infnan) & ~IS_NUMBER_NEG; /* Clear sign for nan. */
1000 else if (flags & PERL_SCAN_TRAILING) {
1001 return numtype | IS_NUMBER_TRAILING;
1008 =for apidoc grok_atou
1010 grok_atou is a safer replacement for atoi and strtol.
1012 grok_atou parses a C-style zero-byte terminated string, looking for
1013 a decimal unsigned integer.
1015 Returns the unsigned integer, if a valid value can be parsed
1016 from the beginning of the string.
1018 Accepts only the decimal digits '0'..'9'.
1020 As opposed to atoi or strtol, grok_atou does NOT allow optional
1021 leading whitespace, or negative inputs. If such features are
1022 required, the calling code needs to explicitly implement those.
1024 If a valid value cannot be parsed, returns either zero (if non-digits
1025 are met before any digits) or UV_MAX (if the value overflows).
1027 Note that extraneous leading zeros also count as an overflow
1028 (meaning that only "0" is the zero).
1030 On failure, the *endptr is also set to NULL, unless endptr is NULL.
1032 Trailing non-digit bytes are allowed if the endptr is non-NULL.
1033 On return the *endptr will contain the pointer to the first non-digit byte.
1035 If the endptr is NULL, the first non-digit byte MUST be
1036 the zero byte terminating the pv, or zero will be returned.
1038 Background: atoi has severe problems with illegal inputs, it cannot be
1039 used for incremental parsing, and therefore should be avoided
1040 atoi and strtol are also affected by locale settings, which can also be
1041 seen as a bug (global state controlled by user environment).
1047 Perl_grok_atou(const char *pv, const char** endptr)
1051 const char* end2; /* Used in case endptr is NULL. */
1052 UV val = 0; /* The return value. */
1054 PERL_ARGS_ASSERT_GROK_ATOU;
1056 eptr = endptr ? endptr : &end2;
1058 /* Single-digit inputs are quite common. */
1061 /* Extra leading zeros cause overflow. */
1066 while (isDIGIT(*s)) {
1067 /* This could be unrolled like in grok_number(), but
1068 * the expected uses of this are not speed-needy, and
1069 * unlikely to need full 64-bitness. */
1070 U8 digit = *s++ - '0';
1071 if (val < uv_max_div_10 ||
1072 (val == uv_max_div_10 && digit <= uv_max_mod_10)) {
1073 val = val * 10 + digit;
1082 *eptr = NULL; /* If no progress, failed to parse anything. */
1085 if (endptr == NULL && *s) {
1086 return 0; /* If endptr is NULL, no trailing non-digits allowed. */
1092 #ifndef USE_QUADMATH
1094 S_mulexp10(NV value, I32 exponent)
1106 /* On OpenVMS VAX we by default use the D_FLOAT double format,
1107 * and that format does not have *easy* capabilities [1] for
1108 * overflowing doubles 'silently' as IEEE fp does. We also need
1109 * to support G_FLOAT on both VAX and Alpha, and though the exponent
1110 * range is much larger than D_FLOAT it still doesn't do silent
1111 * overflow. Therefore we need to detect early whether we would
1112 * overflow (this is the behaviour of the native string-to-float
1113 * conversion routines, and therefore of native applications, too).
1115 * [1] Trying to establish a condition handler to trap floating point
1116 * exceptions is not a good idea. */
1118 /* In UNICOS and in certain Cray models (such as T90) there is no
1119 * IEEE fp, and no way at all from C to catch fp overflows gracefully.
1120 * There is something you can do if you are willing to use some
1121 * inline assembler: the instruction is called DFI-- but that will
1122 * disable *all* floating point interrupts, a little bit too large
1123 * a hammer. Therefore we need to catch potential overflows before
1126 #if ((defined(VMS) && !defined(_IEEE_FP)) || defined(_UNICOS)) && defined(NV_MAX_10_EXP)
1128 const NV exp_v = log10(value);
1129 if (exponent >= NV_MAX_10_EXP || exponent + exp_v >= NV_MAX_10_EXP)
1132 if (-(exponent + exp_v) >= NV_MAX_10_EXP)
1134 while (-exponent >= NV_MAX_10_EXP) {
1135 /* combination does not overflow, but 10^(-exponent) does */
1145 exponent = -exponent;
1146 #ifdef NV_MAX_10_EXP
1147 /* for something like 1234 x 10^-309, the action of calculating
1148 * the intermediate value 10^309 then returning 1234 / (10^309)
1149 * will fail, since 10^309 becomes infinity. In this case try to
1150 * refactor it as 123 / (10^308) etc.
1152 while (value && exponent > NV_MAX_10_EXP) {
1160 #if defined(__osf__)
1161 /* Even with cc -ieee + ieee_set_fp_control(IEEE_TRAP_ENABLE_INV)
1162 * Tru64 fp behavior on inf/nan is somewhat broken. Another way
1163 * to do this would be ieee_set_fp_control(IEEE_TRAP_ENABLE_OVF)
1164 * but that breaks another set of infnan.t tests. */
1165 # define FP_OVERFLOWS_TO_ZERO
1167 for (bit = 1; exponent; bit <<= 1) {
1168 if (exponent & bit) {
1171 #ifdef FP_OVERFLOWS_TO_ZERO
1173 return value < 0 ? -NV_INF : NV_INF;
1175 /* Floating point exceptions are supposed to be turned off,
1176 * but if we're obviously done, don't risk another iteration.
1178 if (exponent == 0) break;
1182 return negative ? value / result : value * result;
1184 #endif /* #ifndef USE_QUADMATH */
1187 Perl_my_atof(pTHX_ const char* s)
1191 Perl_my_atof2(aTHX_ s, &x);
1194 # ifdef USE_LOCALE_NUMERIC
1195 PERL_ARGS_ASSERT_MY_ATOF;
1198 DECLARE_STORE_LC_NUMERIC_SET_TO_NEEDED();
1199 if (PL_numeric_radix_sv && IN_LC(LC_NUMERIC)) {
1200 const char *standard = NULL, *local = NULL;
1201 bool use_standard_radix;
1203 /* Look through the string for the first thing that looks like a
1204 * decimal point: either the value in the current locale or the
1205 * standard fallback of '.'. The one which appears earliest in the
1206 * input string is the one that we should have atof look for. Note
1207 * that we have to determine this beforehand because on some
1208 * systems, Perl_atof2 is just a wrapper around the system's atof.
1210 standard = strchr(s, '.');
1211 local = strstr(s, SvPV_nolen(PL_numeric_radix_sv));
1213 use_standard_radix = standard && (!local || standard < local);
1215 if (use_standard_radix)
1216 SET_NUMERIC_STANDARD();
1220 if (use_standard_radix)
1221 SET_NUMERIC_LOCAL();
1225 RESTORE_LC_NUMERIC();
1235 S_my_atof_infnan(pTHX_ const char* s, bool negative, const char* send, NV* value)
1237 const char *p0 = negative ? s - 1 : s;
1239 int infnan = grok_infnan(&p, send);
1240 if (infnan && p != p0) {
1241 /* If we can generate inf/nan directly, let's do so. */
1243 if ((infnan & IS_NUMBER_INFINITY)) {
1244 *value = (infnan & IS_NUMBER_NEG) ? -NV_INF: NV_INF;
1249 if ((infnan & IS_NUMBER_NAN)) {
1255 /* If still here, we didn't have either NV_INF or NV_NAN,
1256 * and can try falling back to native strtod/strtold.
1258 * (Though, are our NV_INF or NV_NAN ever not defined?)
1260 * The native interface might not recognize all the possible
1261 * inf/nan strings Perl recognizes. What we can try
1262 * is to try faking the input. We will try inf/-inf/nan
1263 * as the most promising/portable input. */
1265 const char* fake = NULL;
1268 if ((infnan & IS_NUMBER_INFINITY)) {
1269 fake = ((infnan & IS_NUMBER_NEG)) ? "-inf" : "inf";
1271 else if ((infnan & IS_NUMBER_NAN)) {
1275 nv = Perl_strtod(fake, &endp);
1277 if ((infnan & IS_NUMBER_INFINITY)) {
1282 /* last resort, may generate SIGFPE */
1283 *value = Perl_exp((NV)1e9);
1284 if ((infnan & IS_NUMBER_NEG))
1287 return (char*)p; /* p, not endp */
1289 else if ((infnan & IS_NUMBER_NAN)) {
1294 /* last resort, may generate SIGFPE */
1295 *value = Perl_log((NV)-1.0);
1297 return (char*)p; /* p, not endp */
1301 #endif /* #ifdef Perl_strtod */
1307 Perl_my_atof2(pTHX_ const char* orig, NV* value)
1309 const char* s = orig;
1310 NV result[3] = {0.0, 0.0, 0.0};
1311 #if defined(USE_PERL_ATOF) || defined(USE_QUADMATH)
1312 const char* send = s + strlen(orig); /* one past the last */
1315 #if defined(USE_PERL_ATOF) && !defined(USE_QUADMATH)
1316 UV accumulator[2] = {0,0}; /* before/after dp */
1317 bool seen_digit = 0;
1318 I32 exp_adjust[2] = {0,0};
1319 I32 exp_acc[2] = {-1, -1};
1320 /* the current exponent adjust for the accumulators */
1325 I32 sig_digits = 0; /* noof significant digits seen so far */
1328 #if defined(USE_PERL_ATOF) || defined(USE_QUADMATH)
1329 PERL_ARGS_ASSERT_MY_ATOF2;
1331 /* leading whitespace */
1348 if ((endp = S_my_atof_infnan(s, negative, send, value)))
1350 result[2] = strtoflt128(s, &endp);
1352 *value = negative ? -result[2] : result[2];
1357 #elif defined(USE_PERL_ATOF)
1359 /* There is no point in processing more significant digits
1360 * than the NV can hold. Note that NV_DIG is a lower-bound value,
1361 * while we need an upper-bound value. We add 2 to account for this;
1362 * since it will have been conservative on both the first and last digit.
1363 * For example a 32-bit mantissa with an exponent of 4 would have
1364 * exact values in the set
1372 * where for the purposes of calculating NV_DIG we would have to discount
1373 * both the first and last digit, since neither can hold all values from
1374 * 0..9; but for calculating the value we must examine those two digits.
1376 #ifdef MAX_SIG_DIG_PLUS
1377 /* It is not necessarily the case that adding 2 to NV_DIG gets all the
1378 possible digits in a NV, especially if NVs are not IEEE compliant
1379 (e.g., long doubles on IRIX) - Allen <allens@cpan.org> */
1380 # define MAX_SIG_DIGITS (NV_DIG+MAX_SIG_DIG_PLUS)
1382 # define MAX_SIG_DIGITS (NV_DIG+2)
1385 /* the max number we can accumulate in a UV, and still safely do 10*N+9 */
1386 #define MAX_ACCUMULATE ( (UV) ((UV_MAX - 9)/10))
1390 if ((endp = S_my_atof_infnan(aTHX_ s, negative, send, value)))
1394 /* we accumulate digits into an integer; when this becomes too
1395 * large, we add the total to NV and start again */
1405 /* don't start counting until we see the first significant
1406 * digit, eg the 5 in 0.00005... */
1407 if (!sig_digits && digit == 0)
1410 if (++sig_digits > MAX_SIG_DIGITS) {
1411 /* limits of precision reached */
1413 ++accumulator[seen_dp];
1414 } else if (digit == 5) {
1415 if (old_digit % 2) { /* round to even - Allen */
1416 ++accumulator[seen_dp];
1424 /* skip remaining digits */
1425 while (isDIGIT(*s)) {
1431 /* warn of loss of precision? */
1434 if (accumulator[seen_dp] > MAX_ACCUMULATE) {
1435 /* add accumulator to result and start again */
1436 result[seen_dp] = S_mulexp10(result[seen_dp],
1438 + (NV)accumulator[seen_dp];
1439 accumulator[seen_dp] = 0;
1440 exp_acc[seen_dp] = 0;
1442 accumulator[seen_dp] = accumulator[seen_dp] * 10 + digit;
1446 else if (!seen_dp && GROK_NUMERIC_RADIX(&s, send)) {
1448 if (sig_digits > MAX_SIG_DIGITS) {
1451 } while (isDIGIT(*s));
1460 result[0] = S_mulexp10(result[0], exp_acc[0]) + (NV)accumulator[0];
1462 result[1] = S_mulexp10(result[1], exp_acc[1]) + (NV)accumulator[1];
1465 if (seen_digit && (isALPHA_FOLD_EQ(*s, 'e'))) {
1466 bool expnegative = 0;
1477 exponent = exponent * 10 + (*s++ - '0');
1479 exponent = -exponent;
1484 /* now apply the exponent */
1487 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0])
1488 + S_mulexp10(result[1],exponent-exp_adjust[1]);
1490 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0]);
1493 /* now apply the sign */
1495 result[2] = -result[2];
1496 #endif /* USE_PERL_ATOF */
1502 =for apidoc isinfnan
1504 Perl_isinfnan() is utility function that returns true if the NV
1505 argument is either an infinity or a NaN, false otherwise. To test
1506 in more detail, use Perl_isinf() and Perl_isnan().
1508 This is also the logical inverse of Perl_isfinite().
1513 Perl_isinfnan(NV nv)
1529 Checks whether the argument would be either an infinity or NaN when used
1530 as a number, but is careful not to trigger non-numeric or uninitialized
1531 warnings. it assumes the caller has done SvGETMAGIC(sv) already.
1537 Perl_isinfnansv(pTHX_ SV *sv)
1539 PERL_ARGS_ASSERT_ISINFNANSV;
1543 return Perl_isinfnan(SvNVX(sv));
1548 const char *s = SvPV_nomg_const(sv, len);
1549 return cBOOL(grok_infnan(&s, s+len));
1554 /* C99 has truncl, pre-C99 Solaris had aintl. We can use either with
1555 * copysignl to emulate modfl, which is in some platforms missing or
1557 # if defined(HAS_TRUNCL) && defined(HAS_COPYSIGNL)
1559 Perl_my_modfl(long double x, long double *ip)
1562 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
1564 # elif defined(HAS_AINTL) && defined(HAS_COPYSIGNL)
1566 Perl_my_modfl(long double x, long double *ip)
1569 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
1574 /* Similarly, with ilogbl and scalbnl we can emulate frexpl. */
1575 #if ! defined(HAS_FREXPL) && defined(HAS_ILOGBL) && defined(HAS_SCALBNL)
1577 Perl_my_frexpl(long double x, int *e) {
1578 *e = x == 0.0L ? 0 : ilogbl(x) + 1;
1579 return (scalbnl(x, -*e));
1584 =for apidoc Perl_signbit
1586 Return a non-zero integer if the sign bit on an NV is set, and 0 if
1589 If Configure detects this system has a signbit() that will work with
1590 our NVs, then we just use it via the #define in perl.h. Otherwise,
1591 fall back on this implementation. The main use of this function
1594 Configure notes: This function is called 'Perl_signbit' instead of a
1595 plain 'signbit' because it is easy to imagine a system having a signbit()
1596 function or macro that doesn't happen to work with our particular choice
1597 of NVs. We shouldn't just re-#define signbit as Perl_signbit and expect
1598 the standard system headers to be happy. Also, this is a no-context
1599 function (no pTHX_) because Perl_signbit() is usually re-#defined in
1600 perl.h as a simple macro call to the system's signbit().
1601 Users should just always call Perl_signbit().
1605 #if !defined(HAS_SIGNBIT)
1607 Perl_signbit(NV x) {
1608 # ifdef Perl_fp_class_nzero
1610 return Perl_fp_class_nzero(x);
1612 return (x < 0.0) ? 1 : 0;
1618 * c-indentation-style: bsd
1620 * indent-tabs-mode: nil
1623 * ex: set ts=8 sts=4 sw=4 et: