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 PERL_ARGS_ASSERT_GROK_NUMERIC_RADIX;
523 #ifdef USE_LOCALE_NUMERIC
525 if (IN_LC(LC_NUMERIC)) {
528 bool matches_radix = FALSE;
529 DECLARATION_FOR_LC_NUMERIC_MANIPULATION;
531 STORE_LC_NUMERIC_FORCE_TO_UNDERLYING();
533 radix = SvPV(PL_numeric_radix_sv, len);
534 radix = savepvn(radix, len);
536 RESTORE_LC_NUMERIC();
538 if (*sp + len <= send) {
539 matches_radix = memEQ(*sp, radix, len);
552 /* always try "." if numeric radix didn't match because
553 * we may have data from different locales mixed */
554 if (*sp < send && **sp == '.') {
563 =for apidoc grok_infnan
565 Helper for C<grok_number()>, accepts various ways of spelling "infinity"
566 or "not a number", and returns one of the following flag combinations:
570 IS_NUMBER_INFINITY | IS_NUMBER_NEG
571 IS_NUMBER_NAN | IS_NUMBER_NEG
574 possibly |-ed with C<IS_NUMBER_TRAILING>.
576 If an infinity or a not-a-number is recognized, C<*sp> will point to
577 one byte past the end of the recognized string. If the recognition fails,
578 zero is returned, and C<*sp> will not move.
584 Perl_grok_infnan(pTHX_ const char** sp, const char* send)
588 #if defined(NV_INF) || defined(NV_NAN)
589 bool odh = FALSE; /* one-dot-hash: 1.#INF */
591 PERL_ARGS_ASSERT_GROK_INFNAN;
594 s++; if (s == send) return 0;
596 else if (*s == '-') {
597 flags |= IS_NUMBER_NEG; /* Yes, -NaN happens. Incorrect but happens. */
598 s++; if (s == send) return 0;
602 /* Visual C: 1.#SNAN, -1.#QNAN, 1#INF, 1.#IND (maybe also 1.#NAN)
603 * Let's keep the dot optional. */
604 s++; if (s == send) return 0;
606 s++; if (s == send) return 0;
609 s++; if (s == send) return 0;
615 if (isALPHA_FOLD_EQ(*s, 'I')) {
616 /* INF or IND (1.#IND is "indeterminate", a certain type of NAN) */
618 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
619 s++; if (s == send) return 0;
620 if (isALPHA_FOLD_EQ(*s, 'F')) {
622 if (s < send && (isALPHA_FOLD_EQ(*s, 'I'))) {
624 flags | IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT | IS_NUMBER_TRAILING;
625 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return fail;
626 s++; if (s == send || isALPHA_FOLD_NE(*s, 'I')) return fail;
627 s++; if (s == send || isALPHA_FOLD_NE(*s, 'T')) return fail;
628 s++; if (s == send || isALPHA_FOLD_NE(*s, 'Y')) return fail;
631 while (*s == '0') { /* 1.#INF00 */
635 while (s < send && isSPACE(*s))
637 if (s < send && *s) {
638 flags |= IS_NUMBER_TRAILING;
640 flags |= IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT;
642 else if (isALPHA_FOLD_EQ(*s, 'D') && odh) { /* 1.#IND */
644 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
645 while (*s == '0') { /* 1.#IND00 */
649 flags |= IS_NUMBER_TRAILING;
655 /* Maybe NAN of some sort */
657 if (isALPHA_FOLD_EQ(*s, 'S') || isALPHA_FOLD_EQ(*s, 'Q')) {
659 /* XXX do something with the snan/qnan difference */
660 s++; if (s == send) return 0;
663 if (isALPHA_FOLD_EQ(*s, 'N')) {
664 s++; if (s == send || isALPHA_FOLD_NE(*s, 'A')) return 0;
665 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
668 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
670 /* NaN can be followed by various stuff (NaNQ, NaNS), but
671 * there are also multiple different NaN values, and some
672 * implementations output the "payload" values,
673 * e.g. NaN123, NAN(abc), while some legacy implementations
674 * have weird stuff like NaN%. */
675 if (isALPHA_FOLD_EQ(*s, 'q') ||
676 isALPHA_FOLD_EQ(*s, 's')) {
677 /* "nanq" or "nans" are ok, though generating
678 * these portably is tricky. */
682 /* C99 style "nan(123)" or Perlish equivalent "nan($uv)". */
686 return flags | IS_NUMBER_TRAILING;
689 while (t < send && *t && *t != ')') {
693 return flags | IS_NUMBER_TRAILING;
698 if (s[0] == '0' && s + 2 < t &&
699 isALPHA_FOLD_EQ(s[1], 'x') &&
702 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES;
703 nanval = grok_hex(s, &len, &flags, NULL);
704 if ((flags & PERL_SCAN_GREATER_THAN_UV_MAX)) {
707 nantype = IS_NUMBER_IN_UV;
710 } else if (s[0] == '0' && s + 2 < t &&
711 isALPHA_FOLD_EQ(s[1], 'b') &&
712 (s[2] == '0' || s[2] == '1')) {
714 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES;
715 nanval = grok_bin(s, &len, &flags, NULL);
716 if ((flags & PERL_SCAN_GREATER_THAN_UV_MAX)) {
719 nantype = IS_NUMBER_IN_UV;
725 grok_number_flags(s, t - s, &nanval,
727 PERL_SCAN_ALLOW_UNDERSCORES);
728 /* Unfortunately grok_number_flags() doesn't
729 * tell how far we got and the ')' will always
730 * be "trailing", so we need to double-check
731 * whether we had something dubious. */
732 for (u = s; u < t; u++) {
734 flags |= IS_NUMBER_TRAILING;
741 /* XXX Doesn't do octal: nan("0123").
742 * Probably not a big loss. */
744 if ((nantype & IS_NUMBER_NOT_INT) ||
745 !(nantype && IS_NUMBER_IN_UV)) {
746 /* XXX the nanval is currently unused, that is,
747 * not inserted as the NaN payload of the NV.
748 * But the above code already parses the C99
749 * nan(...) format. See below, and see also
750 * the nan() in POSIX.xs.
752 * Certain configuration combinations where
753 * NVSIZE is greater than UVSIZE mean that
754 * a single UV cannot contain all the possible
755 * NaN payload bits. There would need to be
756 * some more generic syntax than "nan($uv)".
758 * Issues to keep in mind:
760 * (1) In most common cases there would
761 * not be an integral number of bytes that
762 * could be set, only a certain number of bits.
763 * For example for the common case of
764 * NVSIZE == UVSIZE == 8 there is room for 52
765 * bits in the payload, but the most significant
766 * bit is commonly reserved for the
767 * signaling/quiet bit, leaving 51 bits.
768 * Furthermore, the C99 nan() is supposed
769 * to generate quiet NaNs, so it is doubtful
770 * whether it should be able to generate
771 * signaling NaNs. For the x86 80-bit doubles
772 * (if building a long double Perl) there would
773 * be 62 bits (s/q bit being the 63rd).
775 * (2) Endianness of the payload bits. If the
776 * payload is specified as an UV, the low-order
777 * bits of the UV are naturally little-endianed
778 * (rightmost) bits of the payload. The endianness
779 * of UVs and NVs can be different. */
783 flags |= IS_NUMBER_TRAILING;
786 /* Looked like nan(...), but no close paren. */
787 flags |= IS_NUMBER_TRAILING;
790 while (s < send && isSPACE(*s))
792 if (s < send && *s) {
793 /* Note that we here implicitly accept (parse as
794 * "nan", but with warnings) also any other weird
795 * trailing stuff for "nan". In the above we just
796 * check that if we got the C99-style "nan(...)",
797 * the "..." looks sane.
798 * If in future we accept more ways of specifying
799 * the nan payload, the accepting would happen around
801 flags |= IS_NUMBER_TRAILING;
810 while (s < send && isSPACE(*s))
814 PERL_UNUSED_ARG(send);
815 #endif /* #if defined(NV_INF) || defined(NV_NAN) */
821 =for apidoc grok_number_flags
823 Recognise (or not) a number. The type of the number is returned
824 (0 if unrecognised), otherwise it is a bit-ORed combination of
825 C<IS_NUMBER_IN_UV>, C<IS_NUMBER_GREATER_THAN_UV_MAX>, C<IS_NUMBER_NOT_INT>,
826 C<IS_NUMBER_NEG>, C<IS_NUMBER_INFINITY>, C<IS_NUMBER_NAN> (defined in perl.h).
828 If the value of the number can fit in a UV, it is returned in C<*valuep>.
829 C<IS_NUMBER_IN_UV> will be set to indicate that C<*valuep> is valid, C<IS_NUMBER_IN_UV>
830 will never be set unless C<*valuep> is valid, but C<*valuep> may have been assigned
831 to during processing even though C<IS_NUMBER_IN_UV> is not set on return.
832 If C<valuep> is C<NULL>, C<IS_NUMBER_IN_UV> will be set for the same cases as when
833 C<valuep> is non-C<NULL>, but no actual assignment (or SEGV) will occur.
835 C<IS_NUMBER_NOT_INT> will be set with C<IS_NUMBER_IN_UV> if trailing decimals were
836 seen (in which case C<*valuep> gives the true value truncated to an integer), and
837 C<IS_NUMBER_NEG> if the number is negative (in which case C<*valuep> holds the
838 absolute value). C<IS_NUMBER_IN_UV> is not set if e notation was used or the
839 number is larger than a UV.
841 C<flags> allows only C<PERL_SCAN_TRAILING>, which allows for trailing
842 non-numeric text on an otherwise successful I<grok>, setting
843 C<IS_NUMBER_TRAILING> on the result.
845 =for apidoc grok_number
847 Identical to C<grok_number_flags()> with C<flags> set to zero.
852 Perl_grok_number(pTHX_ const char *pv, STRLEN len, UV *valuep)
854 PERL_ARGS_ASSERT_GROK_NUMBER;
856 return grok_number_flags(pv, len, valuep, 0);
859 static const UV uv_max_div_10 = UV_MAX / 10;
860 static const U8 uv_max_mod_10 = UV_MAX % 10;
863 Perl_grok_number_flags(pTHX_ const char *pv, STRLEN len, UV *valuep, U32 flags)
866 const char * const send = pv + len;
870 PERL_ARGS_ASSERT_GROK_NUMBER_FLAGS;
872 while (s < send && isSPACE(*s))
876 } else if (*s == '-') {
878 numtype = IS_NUMBER_NEG;
886 /* The first digit (after optional sign): note that might
887 * also point to "infinity" or "nan", or "1.#INF". */
890 /* next must be digit or the radix separator or beginning of infinity/nan */
892 /* UVs are at least 32 bits, so the first 9 decimal digits cannot
895 /* This construction seems to be more optimiser friendly.
896 (without it gcc does the isDIGIT test and the *s - '0' separately)
897 With it gcc on arm is managing 6 instructions (6 cycles) per digit.
898 In theory the optimiser could deduce how far to unroll the loop
899 before checking for overflow. */
901 int digit = *s - '0';
902 if (inRANGE(digit, 0, 9)) {
903 value = value * 10 + digit;
906 if (inRANGE(digit, 0, 9)) {
907 value = value * 10 + digit;
910 if (inRANGE(digit, 0, 9)) {
911 value = value * 10 + digit;
914 if (inRANGE(digit, 0, 9)) {
915 value = value * 10 + digit;
918 if (inRANGE(digit, 0, 9)) {
919 value = value * 10 + digit;
922 if (inRANGE(digit, 0, 9)) {
923 value = value * 10 + digit;
926 if (inRANGE(digit, 0, 9)) {
927 value = value * 10 + digit;
930 if (inRANGE(digit, 0, 9)) {
931 value = value * 10 + digit;
933 /* Now got 9 digits, so need to check
934 each time for overflow. */
936 while ( inRANGE(digit, 0, 9)
937 && (value < uv_max_div_10
938 || (value == uv_max_div_10
939 && digit <= uv_max_mod_10))) {
940 value = value * 10 + digit;
946 if (inRANGE(digit, 0, 9)
949 skip the remaining digits, don't
950 worry about setting *valuep. */
953 } while (s < send && isDIGIT(*s));
955 IS_NUMBER_GREATER_THAN_UV_MAX;
975 numtype |= IS_NUMBER_IN_UV;
980 if (GROK_NUMERIC_RADIX(&s, send)) {
981 numtype |= IS_NUMBER_NOT_INT;
982 while (s < send && isDIGIT(*s)) /* optional digits after the radix */
986 else if (GROK_NUMERIC_RADIX(&s, send)) {
987 numtype |= IS_NUMBER_NOT_INT | IS_NUMBER_IN_UV; /* valuep assigned below */
988 /* no digits before the radix means we need digits after it */
989 if (s < send && isDIGIT(*s)) {
992 } while (s < send && isDIGIT(*s));
994 /* integer approximation is valid - it's 0. */
1002 if (s > d && s < send) {
1003 /* we can have an optional exponent part */
1004 if (isALPHA_FOLD_EQ(*s, 'e')) {
1006 if (s < send && (*s == '-' || *s == '+'))
1008 if (s < send && isDIGIT(*s)) {
1011 } while (s < send && isDIGIT(*s));
1013 else if (flags & PERL_SCAN_TRAILING)
1014 return numtype | IS_NUMBER_TRAILING;
1018 /* The only flag we keep is sign. Blow away any "it's UV" */
1019 numtype &= IS_NUMBER_NEG;
1020 numtype |= IS_NUMBER_NOT_INT;
1023 while (s < send && isSPACE(*s))
1027 if (memEQs(pv, len, "0 but true")) {
1030 return IS_NUMBER_IN_UV;
1032 /* We could be e.g. at "Inf" or "NaN", or at the "#" of "1.#INF". */
1033 if ((s + 2 < send) && strchr("inqs#", toFOLD(*s))) {
1034 /* Really detect inf/nan. Start at d, not s, since the above
1035 * code might have already consumed the "1." or "1". */
1036 const int infnan = Perl_grok_infnan(aTHX_ &d, send);
1037 if ((infnan & IS_NUMBER_INFINITY)) {
1038 return (numtype | infnan); /* Keep sign for infinity. */
1040 else if ((infnan & IS_NUMBER_NAN)) {
1041 return (numtype | infnan) & ~IS_NUMBER_NEG; /* Clear sign for nan. */
1044 else if (flags & PERL_SCAN_TRAILING) {
1045 return numtype | IS_NUMBER_TRAILING;
1052 =for apidoc grok_atoUV
1054 parse a string, looking for a decimal unsigned integer.
1056 On entry, C<pv> points to the beginning of the string;
1057 C<valptr> points to a UV that will receive the converted value, if found;
1058 C<endptr> is either NULL or points to a variable that points to one byte
1059 beyond the point in C<pv> that this routine should examine.
1060 If C<endptr> is NULL, C<pv> is assumed to be NUL-terminated.
1062 Returns FALSE if C<pv> doesn't represent a valid unsigned integer value (with
1063 no leading zeros). Otherwise it returns TRUE, and sets C<*valptr> to that
1066 If you constrain the portion of C<pv> that is looked at by this function (by
1067 passing a non-NULL C<endptr>), and if the intial bytes of that portion form a
1068 valid value, it will return TRUE, setting C<*endptr> to the byte following the
1069 final digit of the value. But if there is no constraint at what's looked at,
1070 all of C<pv> must be valid in order for TRUE to be returned.
1072 The only characters this accepts are the decimal digits '0'..'9'.
1074 As opposed to L<atoi(3)> or L<strtol(3)>, C<grok_atoUV> does NOT allow optional
1075 leading whitespace, nor negative inputs. If such features are required, the
1076 calling code needs to explicitly implement those.
1078 Note that this function returns FALSE for inputs that would overflow a UV,
1079 or have leading zeros. Thus a single C<0> is accepted, but not C<00> nor
1080 C<01>, C<002>, I<etc>.
1082 Background: C<atoi> has severe problems with illegal inputs, it cannot be
1083 used for incremental parsing, and therefore should be avoided
1084 C<atoi> and C<strtol> are also affected by locale settings, which can also be
1085 seen as a bug (global state controlled by user environment).
1092 Perl_grok_atoUV(const char *pv, UV *valptr, const char** endptr)
1096 const char* end2; /* Used in case endptr is NULL. */
1097 UV val = 0; /* The parsed value. */
1099 PERL_ARGS_ASSERT_GROK_ATOUV;
1105 end2 = s + strlen(s);
1115 /* Single-digit inputs are quite common. */
1117 if (s < *eptr && isDIGIT(*s)) {
1118 /* Fail on extra leading zeros. */
1121 while (s < *eptr && isDIGIT(*s)) {
1122 /* This could be unrolled like in grok_number(), but
1123 * the expected uses of this are not speed-needy, and
1124 * unlikely to need full 64-bitness. */
1125 const U8 digit = *s++ - '0';
1126 if (val < uv_max_div_10 ||
1127 (val == uv_max_div_10 && digit <= uv_max_mod_10)) {
1128 val = val * 10 + digit;
1135 if (endptr == NULL) {
1137 return FALSE; /* If endptr is NULL, no trailing non-digits allowed. */
1150 S_mulexp10(NV value, I32 exponent)
1162 /* On OpenVMS VAX we by default use the D_FLOAT double format,
1163 * and that format does not have *easy* capabilities [1] for
1164 * overflowing doubles 'silently' as IEEE fp does. We also need
1165 * to support G_FLOAT on both VAX and Alpha, and though the exponent
1166 * range is much larger than D_FLOAT it still doesn't do silent
1167 * overflow. Therefore we need to detect early whether we would
1168 * overflow (this is the behaviour of the native string-to-float
1169 * conversion routines, and therefore of native applications, too).
1171 * [1] Trying to establish a condition handler to trap floating point
1172 * exceptions is not a good idea. */
1174 /* In UNICOS and in certain Cray models (such as T90) there is no
1175 * IEEE fp, and no way at all from C to catch fp overflows gracefully.
1176 * There is something you can do if you are willing to use some
1177 * inline assembler: the instruction is called DFI-- but that will
1178 * disable *all* floating point interrupts, a little bit too large
1179 * a hammer. Therefore we need to catch potential overflows before
1182 #if ((defined(VMS) && !defined(_IEEE_FP)) || defined(_UNICOS) || defined(DOUBLE_IS_VAX_FLOAT)) && defined(NV_MAX_10_EXP)
1184 const NV exp_v = log10(value);
1185 if (exponent >= NV_MAX_10_EXP || exponent + exp_v >= NV_MAX_10_EXP)
1188 if (-(exponent + exp_v) >= NV_MAX_10_EXP)
1190 while (-exponent >= NV_MAX_10_EXP) {
1191 /* combination does not overflow, but 10^(-exponent) does */
1201 exponent = -exponent;
1202 #ifdef NV_MAX_10_EXP
1203 /* for something like 1234 x 10^-309, the action of calculating
1204 * the intermediate value 10^309 then returning 1234 / (10^309)
1205 * will fail, since 10^309 becomes infinity. In this case try to
1206 * refactor it as 123 / (10^308) etc.
1208 while (value && exponent > NV_MAX_10_EXP) {
1216 #if defined(__osf__)
1217 /* Even with cc -ieee + ieee_set_fp_control(IEEE_TRAP_ENABLE_INV)
1218 * Tru64 fp behavior on inf/nan is somewhat broken. Another way
1219 * to do this would be ieee_set_fp_control(IEEE_TRAP_ENABLE_OVF)
1220 * but that breaks another set of infnan.t tests. */
1221 # define FP_OVERFLOWS_TO_ZERO
1223 for (bit = 1; exponent; bit <<= 1) {
1224 if (exponent & bit) {
1227 #ifdef FP_OVERFLOWS_TO_ZERO
1230 return value < 0 ? -NV_INF : NV_INF;
1232 return value < 0 ? -FLT_MAX : FLT_MAX;
1235 /* Floating point exceptions are supposed to be turned off,
1236 * but if we're obviously done, don't risk another iteration.
1238 if (exponent == 0) break;
1242 return negative ? value / result : value * result;
1244 #endif /* #ifndef Perl_strtod */
1247 # define ATOF(s, x) my_atof2(s, &x)
1249 # define ATOF(s, x) Perl_atof2(s, x)
1253 Perl_my_atof(pTHX_ const char* s)
1255 /* 's' must be NUL terminated */
1259 PERL_ARGS_ASSERT_MY_ATOF;
1261 #if ! defined(USE_LOCALE_NUMERIC)
1268 DECLARATION_FOR_LC_NUMERIC_MANIPULATION;
1269 STORE_LC_NUMERIC_SET_TO_NEEDED();
1270 if (! (PL_numeric_radix_sv && IN_LC(LC_NUMERIC))) {
1275 /* Look through the string for the first thing that looks like a
1276 * decimal point: either the value in the current locale or the
1277 * standard fallback of '.'. The one which appears earliest in the
1278 * input string is the one that we should have atof look for. Note
1279 * that we have to determine this beforehand because on some
1280 * systems, Perl_atof2 is just a wrapper around the system's atof.
1282 const char * const standard_pos = strchr(s, '.');
1283 const char * const local_pos
1284 = strstr(s, SvPV_nolen(PL_numeric_radix_sv));
1285 const bool use_standard_radix
1286 = standard_pos && (!local_pos || standard_pos < local_pos);
1288 if (use_standard_radix) {
1289 SET_NUMERIC_STANDARD();
1290 LOCK_LC_NUMERIC_STANDARD();
1295 if (use_standard_radix) {
1296 UNLOCK_LC_NUMERIC_STANDARD();
1297 SET_NUMERIC_UNDERLYING();
1300 RESTORE_LC_NUMERIC();
1308 #if defined(NV_INF) || defined(NV_NAN)
1311 # pragma warning(push)
1312 # pragma warning(disable:4756;disable:4056)
1315 S_my_atof_infnan(pTHX_ const char* s, bool negative, const char* send, NV* value)
1317 const char *p0 = negative ? s - 1 : s;
1319 const int infnan = grok_infnan(&p, send);
1320 if (infnan && p != p0) {
1321 /* If we can generate inf/nan directly, let's do so. */
1323 if ((infnan & IS_NUMBER_INFINITY)) {
1324 *value = (infnan & IS_NUMBER_NEG) ? -NV_INF: NV_INF;
1329 if ((infnan & IS_NUMBER_NAN)) {
1335 /* If still here, we didn't have either NV_INF or NV_NAN,
1336 * and can try falling back to native strtod/strtold.
1338 * The native interface might not recognize all the possible
1339 * inf/nan strings Perl recognizes. What we can try
1340 * is to try faking the input. We will try inf/-inf/nan
1341 * as the most promising/portable input. */
1343 const char* fake = "silence compiler warning";
1347 if ((infnan & IS_NUMBER_INFINITY)) {
1348 fake = ((infnan & IS_NUMBER_NEG)) ? "-inf" : "inf";
1352 if ((infnan & IS_NUMBER_NAN)) {
1356 assert(strNE(fake, "silence compiler warning"));
1357 nv = Perl_strtod(fake, &endp);
1360 if ((infnan & IS_NUMBER_INFINITY)) {
1365 /* last resort, may generate SIGFPE */
1366 *value = Perl_exp((NV)1e9);
1367 if ((infnan & IS_NUMBER_NEG))
1370 return (char*)p; /* p, not endp */
1374 if ((infnan & IS_NUMBER_NAN)) {
1379 /* last resort, may generate SIGFPE */
1380 *value = Perl_log((NV)-1.0);
1382 return (char*)p; /* p, not endp */
1387 #endif /* #ifdef Perl_strtod */
1392 # pragma warning(pop)
1395 #endif /* if defined(NV_INF) || defined(NV_NAN) */
1398 Perl_my_atof2(pTHX_ const char* orig, NV* value)
1400 PERL_ARGS_ASSERT_MY_ATOF2;
1401 return my_atof3(orig, value, 0);
1405 Perl_my_atof3(pTHX_ const char* orig, NV* value, const STRLEN len)
1407 const char* s = orig;
1408 NV result[3] = {0.0, 0.0, 0.0};
1409 #if defined(USE_PERL_ATOF) || defined(Perl_strtod)
1410 const char* send = s + ((len != 0)
1412 : strlen(orig)); /* one past the last */
1415 #if defined(USE_PERL_ATOF) && !defined(Perl_strtod)
1416 UV accumulator[2] = {0,0}; /* before/after dp */
1417 bool seen_digit = 0;
1418 I32 exp_adjust[2] = {0,0};
1419 I32 exp_acc[2] = {-1, -1};
1420 /* the current exponent adjust for the accumulators */
1425 I32 sig_digits = 0; /* noof significant digits seen so far */
1428 #if defined(USE_PERL_ATOF) || defined(Perl_strtod)
1429 PERL_ARGS_ASSERT_MY_ATOF3;
1431 /* leading whitespace */
1432 while (s < send && isSPACE(*s))
1450 if ((endp = S_my_atof_infnan(aTHX_ s, negative, send, value)))
1453 /* If the length is passed in, the input string isn't NUL-terminated,
1454 * and in it turns out the function below assumes it is; therefore we
1455 * create a copy and NUL-terminate that */
1457 Newx(copy, len + 1, char);
1458 Copy(orig, copy, len, char);
1460 s = copy + (s - orig);
1463 result[2] = Perl_strtod(s, &endp);
1465 /* If we created a copy, 'endp' is in terms of that. Convert back to
1468 s = (s - copy) + (char *) orig;
1469 endp = (endp - copy) + (char *) orig;
1474 *value = negative ? -result[2] : result[2];
1479 #elif defined(USE_PERL_ATOF)
1481 /* There is no point in processing more significant digits
1482 * than the NV can hold. Note that NV_DIG is a lower-bound value,
1483 * while we need an upper-bound value. We add 2 to account for this;
1484 * since it will have been conservative on both the first and last digit.
1485 * For example a 32-bit mantissa with an exponent of 4 would have
1486 * exact values in the set
1494 * where for the purposes of calculating NV_DIG we would have to discount
1495 * both the first and last digit, since neither can hold all values from
1496 * 0..9; but for calculating the value we must examine those two digits.
1498 #ifdef MAX_SIG_DIG_PLUS
1499 /* It is not necessarily the case that adding 2 to NV_DIG gets all the
1500 possible digits in a NV, especially if NVs are not IEEE compliant
1501 (e.g., long doubles on IRIX) - Allen <allens@cpan.org> */
1502 # define MAX_SIG_DIGITS (NV_DIG+MAX_SIG_DIG_PLUS)
1504 # define MAX_SIG_DIGITS (NV_DIG+2)
1507 /* the max number we can accumulate in a UV, and still safely do 10*N+9 */
1508 #define MAX_ACCUMULATE ( (UV) ((UV_MAX - 9)/10))
1510 #if defined(NV_INF) || defined(NV_NAN)
1513 if ((endp = S_my_atof_infnan(aTHX_ s, negative, send, value)))
1518 /* we accumulate digits into an integer; when this becomes too
1519 * large, we add the total to NV and start again */
1529 /* don't start counting until we see the first significant
1530 * digit, eg the 5 in 0.00005... */
1531 if (!sig_digits && digit == 0)
1534 if (++sig_digits > MAX_SIG_DIGITS) {
1535 /* limits of precision reached */
1537 ++accumulator[seen_dp];
1538 } else if (digit == 5) {
1539 if (old_digit % 2) { /* round to even - Allen */
1540 ++accumulator[seen_dp];
1548 /* skip remaining digits */
1549 while (s < send && isDIGIT(*s)) {
1555 /* warn of loss of precision? */
1558 if (accumulator[seen_dp] > MAX_ACCUMULATE) {
1559 /* add accumulator to result and start again */
1560 result[seen_dp] = S_mulexp10(result[seen_dp],
1562 + (NV)accumulator[seen_dp];
1563 accumulator[seen_dp] = 0;
1564 exp_acc[seen_dp] = 0;
1566 accumulator[seen_dp] = accumulator[seen_dp] * 10 + digit;
1570 else if (!seen_dp && GROK_NUMERIC_RADIX(&s, send)) {
1572 if (sig_digits > MAX_SIG_DIGITS) {
1573 while (s < send && isDIGIT(*s)) {
1584 result[0] = S_mulexp10(result[0], exp_acc[0]) + (NV)accumulator[0];
1586 result[1] = S_mulexp10(result[1], exp_acc[1]) + (NV)accumulator[1];
1589 if (s < send && seen_digit && (isALPHA_FOLD_EQ(*s, 'e'))) {
1590 bool expnegative = 0;
1600 while (s < send && isDIGIT(*s))
1601 exponent = exponent * 10 + (*s++ - '0');
1603 exponent = -exponent;
1606 /* now apply the exponent */
1609 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0])
1610 + S_mulexp10(result[1],exponent-exp_adjust[1]);
1612 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0]);
1615 /* now apply the sign */
1617 result[2] = -result[2];
1618 #endif /* USE_PERL_ATOF */
1624 =for apidoc isinfnan
1626 C<Perl_isinfnan()> is utility function that returns true if the NV
1627 argument is either an infinity or a C<NaN>, false otherwise. To test
1628 in more detail, use C<Perl_isinf()> and C<Perl_isnan()>.
1630 This is also the logical inverse of Perl_isfinite().
1635 Perl_isinfnan(NV nv)
1637 PERL_UNUSED_ARG(nv);
1652 Checks whether the argument would be either an infinity or C<NaN> when used
1653 as a number, but is careful not to trigger non-numeric or uninitialized
1654 warnings. it assumes the caller has done C<SvGETMAGIC(sv)> already.
1660 Perl_isinfnansv(pTHX_ SV *sv)
1662 PERL_ARGS_ASSERT_ISINFNANSV;
1666 return Perl_isinfnan(SvNVX(sv));
1671 const char *s = SvPV_nomg_const(sv, len);
1672 return cBOOL(grok_infnan(&s, s+len));
1677 /* C99 has truncl, pre-C99 Solaris had aintl. We can use either with
1678 * copysignl to emulate modfl, which is in some platforms missing or
1680 # if defined(HAS_TRUNCL) && defined(HAS_COPYSIGNL)
1682 Perl_my_modfl(long double x, long double *ip)
1685 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
1687 # elif defined(HAS_AINTL) && defined(HAS_COPYSIGNL)
1689 Perl_my_modfl(long double x, long double *ip)
1692 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
1697 /* Similarly, with ilogbl and scalbnl we can emulate frexpl. */
1698 #if ! defined(HAS_FREXPL) && defined(HAS_ILOGBL) && defined(HAS_SCALBNL)
1700 Perl_my_frexpl(long double x, int *e) {
1701 *e = x == 0.0L ? 0 : ilogbl(x) + 1;
1702 return (scalbnl(x, -*e));
1707 =for apidoc Perl_signbit
1709 Return a non-zero integer if the sign bit on an NV is set, and 0 if
1712 If F<Configure> detects this system has a C<signbit()> that will work with
1713 our NVs, then we just use it via the C<#define> in F<perl.h>. Otherwise,
1714 fall back on this implementation. The main use of this function
1715 is catching C<-0.0>.
1717 C<Configure> notes: This function is called C<'Perl_signbit'> instead of a
1718 plain C<'signbit'> because it is easy to imagine a system having a C<signbit()>
1719 function or macro that doesn't happen to work with our particular choice
1720 of NVs. We shouldn't just re-C<#define> C<signbit> as C<Perl_signbit> and expect
1721 the standard system headers to be happy. Also, this is a no-context
1722 function (no C<pTHX_>) because C<Perl_signbit()> is usually re-C<#defined> in
1723 F<perl.h> as a simple macro call to the system's C<signbit()>.
1724 Users should just always call C<Perl_signbit()>.
1728 #if !defined(HAS_SIGNBIT)
1730 Perl_signbit(NV x) {
1731 # ifdef Perl_fp_class_nzero
1732 return Perl_fp_class_nzero(x);
1733 /* Try finding the high byte, and assume it's highest bit
1734 * is the sign. This assumption is probably wrong somewhere. */
1735 # elif defined(USE_LONG_DOUBLE) && LONG_DOUBLEKIND == LONG_DOUBLE_IS_X86_80_BIT_LITTLE_ENDIAN
1736 return (((unsigned char *)&x)[9] & 0x80);
1737 # elif defined(NV_LITTLE_ENDIAN)
1738 /* Note that NVSIZE is sizeof(NV), which would make the below be
1739 * wrong if the end bytes are unused, which happens with the x86
1740 * 80-bit long doubles, which is why take care of that above. */
1741 return (((unsigned char *)&x)[NVSIZE - 1] & 0x80);
1742 # elif defined(NV_BIG_ENDIAN)
1743 return (((unsigned char *)&x)[0] & 0x80);
1745 /* This last resort fallback is wrong for the negative zero. */
1746 return (x < 0.0) ? 1 : 0;
1752 * ex: set ts=8 sts=4 sw=4 et: