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 if (s < send && (isALPHA_FOLD_EQ(*s, 'I'))) {
611 flags | IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT | IS_NUMBER_TRAILING;
612 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return fail;
613 s++; if (s == send || isALPHA_FOLD_NE(*s, 'I')) return fail;
614 s++; if (s == send || isALPHA_FOLD_NE(*s, 'T')) return fail;
615 s++; if (s == send || isALPHA_FOLD_NE(*s, 'Y')) return fail;
618 while (*s == '0') { /* 1.#INF00 */
622 while (s < send && isSPACE(*s))
624 if (s < send && *s) {
625 flags |= IS_NUMBER_TRAILING;
627 flags |= IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT;
629 else if (isALPHA_FOLD_EQ(*s, 'D') && odh) { /* 1.#IND */
631 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
632 while (*s == '0') { /* 1.#IND00 */
636 flags |= IS_NUMBER_TRAILING;
642 /* Maybe NAN of some sort */
644 if (isALPHA_FOLD_EQ(*s, 'S') || isALPHA_FOLD_EQ(*s, 'Q')) {
646 /* XXX do something with the snan/qnan difference */
647 s++; if (s == send) return 0;
650 if (isALPHA_FOLD_EQ(*s, 'N')) {
651 s++; if (s == send || isALPHA_FOLD_NE(*s, 'A')) return 0;
652 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
655 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
657 /* NaN can be followed by various stuff (NaNQ, NaNS), but
658 * there are also multiple different NaN values, and some
659 * implementations output the "payload" values,
660 * e.g. NaN123, NAN(abc), while some legacy implementations
661 * have weird stuff like NaN%. */
662 if (isALPHA_FOLD_EQ(*s, 'q') ||
663 isALPHA_FOLD_EQ(*s, 's')) {
664 /* "nanq" or "nans" are ok, though generating
665 * these portably is tricky. */
669 /* C99 style "nan(123)" or Perlish equivalent "nan($uv)". */
673 return flags | IS_NUMBER_TRAILING;
676 while (t < send && *t && *t != ')') {
680 return flags | IS_NUMBER_TRAILING;
685 if (s[0] == '0' && s + 2 < t &&
686 isALPHA_FOLD_EQ(s[1], 'x') &&
689 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES;
690 nanval = grok_hex(s, &len, &flags, NULL);
691 if ((flags & PERL_SCAN_GREATER_THAN_UV_MAX)) {
694 nantype = IS_NUMBER_IN_UV;
697 } else if (s[0] == '0' && s + 2 < t &&
698 isALPHA_FOLD_EQ(s[1], 'b') &&
699 (s[2] == '0' || s[2] == '1')) {
701 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES;
702 nanval = grok_bin(s, &len, &flags, NULL);
703 if ((flags & PERL_SCAN_GREATER_THAN_UV_MAX)) {
706 nantype = IS_NUMBER_IN_UV;
712 grok_number_flags(s, t - s, &nanval,
714 PERL_SCAN_ALLOW_UNDERSCORES);
715 /* Unfortunately grok_number_flags() doesn't
716 * tell how far we got and the ')' will always
717 * be "trailing", so we need to double-check
718 * whether we had something dubious. */
719 for (u = s; u < t; u++) {
721 flags |= IS_NUMBER_TRAILING;
728 /* XXX Doesn't do octal: nan("0123").
729 * Probably not a big loss. */
731 if ((nantype & IS_NUMBER_NOT_INT) ||
732 !(nantype && IS_NUMBER_IN_UV)) {
733 /* XXX the nanval is currently unused, that is,
734 * not inserted as the NaN payload of the NV.
735 * But the above code already parses the C99
736 * nan(...) format. See below, and see also
737 * the nan() in POSIX.xs.
739 * Certain configuration combinations where
740 * NVSIZE is greater than UVSIZE mean that
741 * a single UV cannot contain all the possible
742 * NaN payload bits. There would need to be
743 * some more generic syntax than "nan($uv)".
745 * Issues to keep in mind:
747 * (1) In most common cases there would
748 * not be an integral number of bytes that
749 * could be set, only a certain number of bits.
750 * For example for the common case of
751 * NVSIZE == UVSIZE == 8 there is room for 52
752 * bits in the payload, but the most significant
753 * bit is commonly reserved for the
754 * signaling/quiet bit, leaving 51 bits.
755 * Furthermore, the C99 nan() is supposed
756 * to generate quiet NaNs, so it is doubtful
757 * whether it should be able to generate
758 * signaling NaNs. For the x86 80-bit doubles
759 * (if building a long double Perl) there would
760 * be 62 bits (s/q bit being the 63rd).
762 * (2) Endianness of the payload bits. If the
763 * payload is specified as an UV, the low-order
764 * bits of the UV are naturally little-endianed
765 * (rightmost) bits of the payload. The endianness
766 * of UVs and NVs can be different. */
770 flags |= IS_NUMBER_TRAILING;
773 /* Looked like nan(...), but no close paren. */
774 flags |= IS_NUMBER_TRAILING;
777 while (s < send && isSPACE(*s))
779 if (s < send && *s) {
780 /* Note that we here implicitly accept (parse as
781 * "nan", but with warnings) also any other weird
782 * trailing stuff for "nan". In the above we just
783 * check that if we got the C99-style "nan(...)",
784 * the "..." looks sane.
785 * If in future we accept more ways of specifying
786 * the nan payload, the accepting would happen around
788 flags |= IS_NUMBER_TRAILING;
797 while (s < send && isSPACE(*s))
805 =for apidoc grok_number_flags
807 Recognise (or not) a number. The type of the number is returned
808 (0 if unrecognised), otherwise it is a bit-ORed combination of
809 IS_NUMBER_IN_UV, IS_NUMBER_GREATER_THAN_UV_MAX, IS_NUMBER_NOT_INT,
810 IS_NUMBER_NEG, IS_NUMBER_INFINITY, IS_NUMBER_NAN (defined in perl.h).
812 If the value of the number can fit in a UV, it is returned in the *valuep
813 IS_NUMBER_IN_UV will be set to indicate that *valuep is valid, IS_NUMBER_IN_UV
814 will never be set unless *valuep is valid, but *valuep may have been assigned
815 to during processing even though IS_NUMBER_IN_UV is not set on return.
816 If valuep is NULL, IS_NUMBER_IN_UV will be set for the same cases as when
817 valuep is non-NULL, but no actual assignment (or SEGV) will occur.
819 IS_NUMBER_NOT_INT will be set with IS_NUMBER_IN_UV if trailing decimals were
820 seen (in which case *valuep gives the true value truncated to an integer), and
821 IS_NUMBER_NEG if the number is negative (in which case *valuep holds the
822 absolute value). IS_NUMBER_IN_UV is not set if e notation was used or the
823 number is larger than a UV.
825 C<flags> allows only C<PERL_SCAN_TRAILING>, which allows for trailing
826 non-numeric text on an otherwise successful I<grok>, setting
827 C<IS_NUMBER_TRAILING> on the result.
829 =for apidoc grok_number
831 Identical to grok_number_flags() with flags set to zero.
836 Perl_grok_number(pTHX_ const char *pv, STRLEN len, UV *valuep)
838 PERL_ARGS_ASSERT_GROK_NUMBER;
840 return grok_number_flags(pv, len, valuep, 0);
843 static const UV uv_max_div_10 = UV_MAX / 10;
844 static const U8 uv_max_mod_10 = UV_MAX % 10;
847 Perl_grok_number_flags(pTHX_ const char *pv, STRLEN len, UV *valuep, U32 flags)
850 const char * const send = pv + len;
854 PERL_ARGS_ASSERT_GROK_NUMBER_FLAGS;
856 while (s < send && isSPACE(*s))
860 } else if (*s == '-') {
862 numtype = IS_NUMBER_NEG;
870 /* The first digit (after optional sign): note that might
871 * also point to "infinity" or "nan", or "1.#INF". */
874 /* next must be digit or the radix separator or beginning of infinity/nan */
876 /* UVs are at least 32 bits, so the first 9 decimal digits cannot
879 /* This construction seems to be more optimiser friendly.
880 (without it gcc does the isDIGIT test and the *s - '0' separately)
881 With it gcc on arm is managing 6 instructions (6 cycles) per digit.
882 In theory the optimiser could deduce how far to unroll the loop
883 before checking for overflow. */
885 int digit = *s - '0';
886 if (digit >= 0 && digit <= 9) {
887 value = value * 10 + digit;
890 if (digit >= 0 && digit <= 9) {
891 value = value * 10 + digit;
894 if (digit >= 0 && digit <= 9) {
895 value = value * 10 + digit;
898 if (digit >= 0 && digit <= 9) {
899 value = value * 10 + digit;
902 if (digit >= 0 && digit <= 9) {
903 value = value * 10 + digit;
906 if (digit >= 0 && digit <= 9) {
907 value = value * 10 + digit;
910 if (digit >= 0 && digit <= 9) {
911 value = value * 10 + digit;
914 if (digit >= 0 && digit <= 9) {
915 value = value * 10 + digit;
917 /* Now got 9 digits, so need to check
918 each time for overflow. */
920 while (digit >= 0 && digit <= 9
921 && (value < uv_max_div_10
922 || (value == uv_max_div_10
923 && digit <= uv_max_mod_10))) {
924 value = value * 10 + digit;
930 if (digit >= 0 && digit <= 9
933 skip the remaining digits, don't
934 worry about setting *valuep. */
937 } while (s < send && isDIGIT(*s));
939 IS_NUMBER_GREATER_THAN_UV_MAX;
959 numtype |= IS_NUMBER_IN_UV;
964 if (GROK_NUMERIC_RADIX(&s, send)) {
965 numtype |= IS_NUMBER_NOT_INT;
966 while (s < send && isDIGIT(*s)) /* optional digits after the radix */
970 else if (GROK_NUMERIC_RADIX(&s, send)) {
971 numtype |= IS_NUMBER_NOT_INT | IS_NUMBER_IN_UV; /* valuep assigned below */
972 /* no digits before the radix means we need digits after it */
973 if (s < send && isDIGIT(*s)) {
976 } while (s < send && isDIGIT(*s));
978 /* integer approximation is valid - it's 0. */
986 if (s > d && s < send) {
987 /* we can have an optional exponent part */
988 if (isALPHA_FOLD_EQ(*s, 'e')) {
990 if (s < send && (*s == '-' || *s == '+'))
992 if (s < send && isDIGIT(*s)) {
995 } while (s < send && isDIGIT(*s));
997 else if (flags & PERL_SCAN_TRAILING)
998 return numtype | IS_NUMBER_TRAILING;
1002 /* The only flag we keep is sign. Blow away any "it's UV" */
1003 numtype &= IS_NUMBER_NEG;
1004 numtype |= IS_NUMBER_NOT_INT;
1007 while (s < send && isSPACE(*s))
1011 if (len == 10 && memEQ(pv, "0 but true", 10)) {
1014 return IS_NUMBER_IN_UV;
1016 /* We could be e.g. at "Inf" or "NaN", or at the "#" of "1.#INF". */
1017 if ((s + 2 < send) && strchr("inqs#", toFOLD(*s))) {
1018 /* Really detect inf/nan. Start at d, not s, since the above
1019 * code might have already consumed the "1." or "1". */
1020 int infnan = Perl_grok_infnan(aTHX_ &d, send);
1021 if ((infnan & IS_NUMBER_INFINITY)) {
1022 return (numtype | infnan); /* Keep sign for infinity. */
1024 else if ((infnan & IS_NUMBER_NAN)) {
1025 return (numtype | infnan) & ~IS_NUMBER_NEG; /* Clear sign for nan. */
1028 else if (flags & PERL_SCAN_TRAILING) {
1029 return numtype | IS_NUMBER_TRAILING;
1036 =for apidoc grok_atou
1038 grok_atou is a safer replacement for atoi and strtol.
1040 grok_atou parses a C-style zero-byte terminated string, looking for
1041 a decimal unsigned integer.
1043 Returns the unsigned integer, if a valid value can be parsed
1044 from the beginning of the string.
1046 Accepts only the decimal digits '0'..'9'.
1048 As opposed to atoi or strtol, grok_atou does NOT allow optional
1049 leading whitespace, or negative inputs. If such features are
1050 required, the calling code needs to explicitly implement those.
1052 If a valid value cannot be parsed, returns either zero (if non-digits
1053 are met before any digits) or UV_MAX (if the value overflows).
1055 Note that extraneous leading zeros also count as an overflow
1056 (meaning that only "0" is the zero).
1058 On failure, the *endptr is also set to NULL, unless endptr is NULL.
1060 Trailing non-digit bytes are allowed if the endptr is non-NULL.
1061 On return the *endptr will contain the pointer to the first non-digit byte.
1063 If the endptr is NULL, the first non-digit byte MUST be
1064 the zero byte terminating the pv, or zero will be returned.
1066 Background: atoi has severe problems with illegal inputs, it cannot be
1067 used for incremental parsing, and therefore should be avoided
1068 atoi and strtol are also affected by locale settings, which can also be
1069 seen as a bug (global state controlled by user environment).
1075 Perl_grok_atou(const char *pv, const char** endptr)
1079 const char* end2; /* Used in case endptr is NULL. */
1080 UV val = 0; /* The return value. */
1082 PERL_ARGS_ASSERT_GROK_ATOU;
1084 eptr = endptr ? endptr : &end2;
1086 /* Single-digit inputs are quite common. */
1089 /* Extra leading zeros cause overflow. */
1094 while (isDIGIT(*s)) {
1095 /* This could be unrolled like in grok_number(), but
1096 * the expected uses of this are not speed-needy, and
1097 * unlikely to need full 64-bitness. */
1098 U8 digit = *s++ - '0';
1099 if (val < uv_max_div_10 ||
1100 (val == uv_max_div_10 && digit <= uv_max_mod_10)) {
1101 val = val * 10 + digit;
1110 *eptr = NULL; /* If no progress, failed to parse anything. */
1113 if (endptr == NULL && *s) {
1114 return 0; /* If endptr is NULL, no trailing non-digits allowed. */
1120 #ifndef USE_QUADMATH
1122 S_mulexp10(NV value, I32 exponent)
1134 /* On OpenVMS VAX we by default use the D_FLOAT double format,
1135 * and that format does not have *easy* capabilities [1] for
1136 * overflowing doubles 'silently' as IEEE fp does. We also need
1137 * to support G_FLOAT on both VAX and Alpha, and though the exponent
1138 * range is much larger than D_FLOAT it still doesn't do silent
1139 * overflow. Therefore we need to detect early whether we would
1140 * overflow (this is the behaviour of the native string-to-float
1141 * conversion routines, and therefore of native applications, too).
1143 * [1] Trying to establish a condition handler to trap floating point
1144 * exceptions is not a good idea. */
1146 /* In UNICOS and in certain Cray models (such as T90) there is no
1147 * IEEE fp, and no way at all from C to catch fp overflows gracefully.
1148 * There is something you can do if you are willing to use some
1149 * inline assembler: the instruction is called DFI-- but that will
1150 * disable *all* floating point interrupts, a little bit too large
1151 * a hammer. Therefore we need to catch potential overflows before
1154 #if ((defined(VMS) && !defined(_IEEE_FP)) || defined(_UNICOS)) && defined(NV_MAX_10_EXP)
1156 const NV exp_v = log10(value);
1157 if (exponent >= NV_MAX_10_EXP || exponent + exp_v >= NV_MAX_10_EXP)
1160 if (-(exponent + exp_v) >= NV_MAX_10_EXP)
1162 while (-exponent >= NV_MAX_10_EXP) {
1163 /* combination does not overflow, but 10^(-exponent) does */
1173 exponent = -exponent;
1174 #ifdef NV_MAX_10_EXP
1175 /* for something like 1234 x 10^-309, the action of calculating
1176 * the intermediate value 10^309 then returning 1234 / (10^309)
1177 * will fail, since 10^309 becomes infinity. In this case try to
1178 * refactor it as 123 / (10^308) etc.
1180 while (value && exponent > NV_MAX_10_EXP) {
1188 #if defined(__osf__)
1189 /* Even with cc -ieee + ieee_set_fp_control(IEEE_TRAP_ENABLE_INV)
1190 * Tru64 fp behavior on inf/nan is somewhat broken. Another way
1191 * to do this would be ieee_set_fp_control(IEEE_TRAP_ENABLE_OVF)
1192 * but that breaks another set of infnan.t tests. */
1193 # define FP_OVERFLOWS_TO_ZERO
1195 for (bit = 1; exponent; bit <<= 1) {
1196 if (exponent & bit) {
1199 #ifdef FP_OVERFLOWS_TO_ZERO
1201 return value < 0 ? -NV_INF : NV_INF;
1203 /* Floating point exceptions are supposed to be turned off,
1204 * but if we're obviously done, don't risk another iteration.
1206 if (exponent == 0) break;
1210 return negative ? value / result : value * result;
1212 #endif /* #ifndef USE_QUADMATH */
1215 Perl_my_atof(pTHX_ const char* s)
1219 Perl_my_atof2(aTHX_ s, &x);
1222 # ifdef USE_LOCALE_NUMERIC
1223 PERL_ARGS_ASSERT_MY_ATOF;
1226 DECLARE_STORE_LC_NUMERIC_SET_TO_NEEDED();
1227 if (PL_numeric_radix_sv && IN_LC(LC_NUMERIC)) {
1228 const char *standard = NULL, *local = NULL;
1229 bool use_standard_radix;
1231 /* Look through the string for the first thing that looks like a
1232 * decimal point: either the value in the current locale or the
1233 * standard fallback of '.'. The one which appears earliest in the
1234 * input string is the one that we should have atof look for. Note
1235 * that we have to determine this beforehand because on some
1236 * systems, Perl_atof2 is just a wrapper around the system's atof.
1238 standard = strchr(s, '.');
1239 local = strstr(s, SvPV_nolen(PL_numeric_radix_sv));
1241 use_standard_radix = standard && (!local || standard < local);
1243 if (use_standard_radix)
1244 SET_NUMERIC_STANDARD();
1248 if (use_standard_radix)
1249 SET_NUMERIC_LOCAL();
1253 RESTORE_LC_NUMERIC();
1264 # pragma warning(push)
1265 # pragma warning(disable:4756;disable:4056)
1268 S_my_atof_infnan(pTHX_ const char* s, bool negative, const char* send, NV* value)
1270 const char *p0 = negative ? s - 1 : s;
1272 int infnan = grok_infnan(&p, send);
1273 if (infnan && p != p0) {
1274 /* If we can generate inf/nan directly, let's do so. */
1276 if ((infnan & IS_NUMBER_INFINITY)) {
1277 *value = (infnan & IS_NUMBER_NEG) ? -NV_INF: NV_INF;
1282 if ((infnan & IS_NUMBER_NAN)) {
1288 /* If still here, we didn't have either NV_INF or NV_NAN,
1289 * and can try falling back to native strtod/strtold.
1291 * (Though, are our NV_INF or NV_NAN ever not defined?)
1293 * The native interface might not recognize all the possible
1294 * inf/nan strings Perl recognizes. What we can try
1295 * is to try faking the input. We will try inf/-inf/nan
1296 * as the most promising/portable input. */
1298 const char* fake = NULL;
1301 if ((infnan & IS_NUMBER_INFINITY)) {
1302 fake = ((infnan & IS_NUMBER_NEG)) ? "-inf" : "inf";
1304 else if ((infnan & IS_NUMBER_NAN)) {
1308 nv = Perl_strtod(fake, &endp);
1310 if ((infnan & IS_NUMBER_INFINITY)) {
1315 /* last resort, may generate SIGFPE */
1316 *value = Perl_exp((NV)1e9);
1317 if ((infnan & IS_NUMBER_NEG))
1320 return (char*)p; /* p, not endp */
1322 else if ((infnan & IS_NUMBER_NAN)) {
1327 /* last resort, may generate SIGFPE */
1328 *value = Perl_log((NV)-1.0);
1330 return (char*)p; /* p, not endp */
1334 #endif /* #ifdef Perl_strtod */
1339 # pragma warning(pop)
1343 Perl_my_atof2(pTHX_ const char* orig, NV* value)
1345 const char* s = orig;
1346 NV result[3] = {0.0, 0.0, 0.0};
1347 #if defined(USE_PERL_ATOF) || defined(USE_QUADMATH)
1348 const char* send = s + strlen(orig); /* one past the last */
1351 #if defined(USE_PERL_ATOF) && !defined(USE_QUADMATH)
1352 UV accumulator[2] = {0,0}; /* before/after dp */
1353 bool seen_digit = 0;
1354 I32 exp_adjust[2] = {0,0};
1355 I32 exp_acc[2] = {-1, -1};
1356 /* the current exponent adjust for the accumulators */
1361 I32 sig_digits = 0; /* noof significant digits seen so far */
1364 #if defined(USE_PERL_ATOF) || defined(USE_QUADMATH)
1365 PERL_ARGS_ASSERT_MY_ATOF2;
1367 /* leading whitespace */
1384 if ((endp = S_my_atof_infnan(s, negative, send, value)))
1386 result[2] = strtoflt128(s, &endp);
1388 *value = negative ? -result[2] : result[2];
1393 #elif defined(USE_PERL_ATOF)
1395 /* There is no point in processing more significant digits
1396 * than the NV can hold. Note that NV_DIG is a lower-bound value,
1397 * while we need an upper-bound value. We add 2 to account for this;
1398 * since it will have been conservative on both the first and last digit.
1399 * For example a 32-bit mantissa with an exponent of 4 would have
1400 * exact values in the set
1408 * where for the purposes of calculating NV_DIG we would have to discount
1409 * both the first and last digit, since neither can hold all values from
1410 * 0..9; but for calculating the value we must examine those two digits.
1412 #ifdef MAX_SIG_DIG_PLUS
1413 /* It is not necessarily the case that adding 2 to NV_DIG gets all the
1414 possible digits in a NV, especially if NVs are not IEEE compliant
1415 (e.g., long doubles on IRIX) - Allen <allens@cpan.org> */
1416 # define MAX_SIG_DIGITS (NV_DIG+MAX_SIG_DIG_PLUS)
1418 # define MAX_SIG_DIGITS (NV_DIG+2)
1421 /* the max number we can accumulate in a UV, and still safely do 10*N+9 */
1422 #define MAX_ACCUMULATE ( (UV) ((UV_MAX - 9)/10))
1426 if ((endp = S_my_atof_infnan(aTHX_ s, negative, send, value)))
1430 /* we accumulate digits into an integer; when this becomes too
1431 * large, we add the total to NV and start again */
1441 /* don't start counting until we see the first significant
1442 * digit, eg the 5 in 0.00005... */
1443 if (!sig_digits && digit == 0)
1446 if (++sig_digits > MAX_SIG_DIGITS) {
1447 /* limits of precision reached */
1449 ++accumulator[seen_dp];
1450 } else if (digit == 5) {
1451 if (old_digit % 2) { /* round to even - Allen */
1452 ++accumulator[seen_dp];
1460 /* skip remaining digits */
1461 while (isDIGIT(*s)) {
1467 /* warn of loss of precision? */
1470 if (accumulator[seen_dp] > MAX_ACCUMULATE) {
1471 /* add accumulator to result and start again */
1472 result[seen_dp] = S_mulexp10(result[seen_dp],
1474 + (NV)accumulator[seen_dp];
1475 accumulator[seen_dp] = 0;
1476 exp_acc[seen_dp] = 0;
1478 accumulator[seen_dp] = accumulator[seen_dp] * 10 + digit;
1482 else if (!seen_dp && GROK_NUMERIC_RADIX(&s, send)) {
1484 if (sig_digits > MAX_SIG_DIGITS) {
1487 } while (isDIGIT(*s));
1496 result[0] = S_mulexp10(result[0], exp_acc[0]) + (NV)accumulator[0];
1498 result[1] = S_mulexp10(result[1], exp_acc[1]) + (NV)accumulator[1];
1501 if (seen_digit && (isALPHA_FOLD_EQ(*s, 'e'))) {
1502 bool expnegative = 0;
1513 exponent = exponent * 10 + (*s++ - '0');
1515 exponent = -exponent;
1520 /* now apply the exponent */
1523 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0])
1524 + S_mulexp10(result[1],exponent-exp_adjust[1]);
1526 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0]);
1529 /* now apply the sign */
1531 result[2] = -result[2];
1532 #endif /* USE_PERL_ATOF */
1538 =for apidoc isinfnan
1540 Perl_isinfnan() is utility function that returns true if the NV
1541 argument is either an infinity or a NaN, false otherwise. To test
1542 in more detail, use Perl_isinf() and Perl_isnan().
1544 This is also the logical inverse of Perl_isfinite().
1549 Perl_isinfnan(NV nv)
1565 Checks whether the argument would be either an infinity or NaN when used
1566 as a number, but is careful not to trigger non-numeric or uninitialized
1567 warnings. it assumes the caller has done SvGETMAGIC(sv) already.
1573 Perl_isinfnansv(pTHX_ SV *sv)
1575 PERL_ARGS_ASSERT_ISINFNANSV;
1579 return Perl_isinfnan(SvNVX(sv));
1584 const char *s = SvPV_nomg_const(sv, len);
1585 return cBOOL(grok_infnan(&s, s+len));
1590 /* C99 has truncl, pre-C99 Solaris had aintl. We can use either with
1591 * copysignl to emulate modfl, which is in some platforms missing or
1593 # if defined(HAS_TRUNCL) && defined(HAS_COPYSIGNL)
1595 Perl_my_modfl(long double x, long double *ip)
1598 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
1600 # elif defined(HAS_AINTL) && defined(HAS_COPYSIGNL)
1602 Perl_my_modfl(long double x, long double *ip)
1605 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
1610 /* Similarly, with ilogbl and scalbnl we can emulate frexpl. */
1611 #if ! defined(HAS_FREXPL) && defined(HAS_ILOGBL) && defined(HAS_SCALBNL)
1613 Perl_my_frexpl(long double x, int *e) {
1614 *e = x == 0.0L ? 0 : ilogbl(x) + 1;
1615 return (scalbnl(x, -*e));
1620 =for apidoc Perl_signbit
1622 Return a non-zero integer if the sign bit on an NV is set, and 0 if
1625 If Configure detects this system has a signbit() that will work with
1626 our NVs, then we just use it via the #define in perl.h. Otherwise,
1627 fall back on this implementation. The main use of this function
1630 Configure notes: This function is called 'Perl_signbit' instead of a
1631 plain 'signbit' because it is easy to imagine a system having a signbit()
1632 function or macro that doesn't happen to work with our particular choice
1633 of NVs. We shouldn't just re-#define signbit as Perl_signbit and expect
1634 the standard system headers to be happy. Also, this is a no-context
1635 function (no pTHX_) because Perl_signbit() is usually re-#defined in
1636 perl.h as a simple macro call to the system's signbit().
1637 Users should just always call Perl_signbit().
1641 #if !defined(HAS_SIGNBIT)
1643 Perl_signbit(NV x) {
1644 # ifdef Perl_fp_class_nzero
1646 return Perl_fp_class_nzero(x);
1648 return (x < 0.0) ? 1 : 0;
1654 * c-indentation-style: bsd
1656 * indent-tabs-mode: nil
1659 * ex: set ts=8 sts=4 sw=4 et: