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 that are still 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_number_flags
553 Recognise (or not) a number. The type of the number is returned
554 (0 if unrecognised), otherwise it is a bit-ORed combination of
555 IS_NUMBER_IN_UV, IS_NUMBER_GREATER_THAN_UV_MAX, IS_NUMBER_NOT_INT,
556 IS_NUMBER_NEG, IS_NUMBER_INFINITY, IS_NUMBER_NAN (defined in perl.h).
558 If the value of the number can fit in a UV, it is returned in the *valuep
559 IS_NUMBER_IN_UV will be set to indicate that *valuep is valid, IS_NUMBER_IN_UV
560 will never be set unless *valuep is valid, but *valuep may have been assigned
561 to during processing even though IS_NUMBER_IN_UV is not set on return.
562 If valuep is NULL, IS_NUMBER_IN_UV will be set for the same cases as when
563 valuep is non-NULL, but no actual assignment (or SEGV) will occur.
565 IS_NUMBER_NOT_INT will be set with IS_NUMBER_IN_UV if trailing decimals were
566 seen (in which case *valuep gives the true value truncated to an integer), and
567 IS_NUMBER_NEG if the number is negative (in which case *valuep holds the
568 absolute value). IS_NUMBER_IN_UV is not set if e notation was used or the
569 number is larger than a UV.
571 C<flags> allows only C<PERL_SCAN_TRAILING>, which allows for trailing
572 non-numeric text on an otherwise successful I<grok>, setting
573 C<IS_NUMBER_TRAILING> on the result.
575 =for apidoc grok_number
577 Identical to grok_number_flags() with flags set to zero.
582 Perl_grok_number(pTHX_ const char *pv, STRLEN len, UV *valuep)
584 PERL_ARGS_ASSERT_GROK_NUMBER;
586 return grok_number_flags(pv, len, valuep, 0);
589 /* Peek ahead to see whether this could be Inf/NaN/qNaN/snan/1.#INF */
590 #define INFNAN_PEEK(s, send) \
592 ((isALPHA_FOLD_EQ(*s, 'I') || isALPHA_FOLD_EQ(*s, 'N')) || \
594 (isALPHA_FOLD_EQ(*s, 'Q') || isALPHA_FOLD_EQ(*s, 'S')) && \
595 isALPHA_FOLD_EQ(s[1], 'N')) || \
597 (*s == '1' && ((s[1] == '.' && s[2] == '#') || s[1] == '#')))))
600 =for apidoc grok_infnan
602 Helper for grok_number(), accepts various ways of spelling "infinity"
603 or "not a number", and returns one of the following flag combinations:
607 IS_NUMBER_INFINITE | IS_NUMBER_NEG
608 IS_NUMBER_NAN | IS_NUMBER_NEG
611 If an infinity or not-a-number is recognized, the *sp will point to
612 one past the end of the recognized string. If the recognition fails,
613 zero is returned, and the *sp will not move.
619 Perl_grok_infnan(const char** sp, const char* send)
624 PERL_ARGS_ASSERT_GROK_INFNAN;
627 flags |= IS_NUMBER_NEG; /* Yes, -NaN happens. Incorrect but happens. */
628 s++; if (s == send) return 0;
632 /* Visual C: 1.#SNAN, -1.#QNAN, 1#INF, 1#.IND (maybe also 1.#NAN) */
633 s++; if (s == send) return 0;
635 s++; if (s == send) return 0;
638 s++; if (s == send) return 0;
643 if (isALPHA_FOLD_EQ(*s, 'I')) {
644 /* INF or IND (1.#IND is indeterminate, a certain type of NAN) */
645 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
646 s++; if (s == send) return 0;
647 if (isALPHA_FOLD_EQ(*s, 'F')) {
649 if (s < send && (isALPHA_FOLD_EQ(*s, 'I'))) {
650 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
651 s++; if (s == send || isALPHA_FOLD_NE(*s, 'I')) return 0;
652 s++; if (s == send || isALPHA_FOLD_NE(*s, 'T')) return 0;
653 /* XXX maybe also grok "infinite"? */
654 s++; if (s == send || isALPHA_FOLD_NE(*s, 'Y')) return 0;
658 flags |= IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT;
660 else if (isALPHA_FOLD_EQ(*s, 'D')) {
662 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
671 if (isALPHA_FOLD_EQ(*s, 'S') || isALPHA_FOLD_EQ(*s, 'Q')) {
673 /* XXX do something with the snan/qnan difference */
674 s++; if (s == send) return 0;
677 if (isALPHA_FOLD_EQ(*s, 'N')) {
678 s++; if (s == send || isALPHA_FOLD_NE(*s, 'A')) return 0;
679 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0;
682 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
684 /* NaN can be followed by various stuff since there are
685 * multiple different NaN values, and some implementations
686 * output the "payload" values, e.g. NaN123, NAN(abc),
687 * some implementation just have weird stuff like NaN%. */
700 static const UV uv_max_div_10 = UV_MAX / 10;
701 static const U8 uv_max_mod_10 = UV_MAX % 10;
704 Perl_grok_number_flags(pTHX_ const char *pv, STRLEN len, UV *valuep, U32 flags)
707 const char * const send = pv + len;
713 PERL_ARGS_ASSERT_GROK_NUMBER_FLAGS;
715 while (s < send && isSPACE(*s))
719 } else if (*s == '-') {
721 numtype = IS_NUMBER_NEG;
729 /* The first digit (after optional sign): note that might
730 * also point to "infinity" or "nan". */
733 /* next must be digit or the radix separator or beginning of infinity */
735 /* UVs are at least 32 bits, so the first 9 decimal digits cannot
738 /* This construction seems to be more optimiser friendly.
739 (without it gcc does the isDIGIT test and the *s - '0' separately)
740 With it gcc on arm is managing 6 instructions (6 cycles) per digit.
741 In theory the optimiser could deduce how far to unroll the loop
742 before checking for overflow. */
744 int digit = *s - '0';
745 if (digit >= 0 && digit <= 9) {
746 value = value * 10 + digit;
749 if (digit >= 0 && digit <= 9) {
750 value = value * 10 + digit;
753 if (digit >= 0 && digit <= 9) {
754 value = value * 10 + digit;
757 if (digit >= 0 && digit <= 9) {
758 value = value * 10 + digit;
761 if (digit >= 0 && digit <= 9) {
762 value = value * 10 + digit;
765 if (digit >= 0 && digit <= 9) {
766 value = value * 10 + digit;
769 if (digit >= 0 && digit <= 9) {
770 value = value * 10 + digit;
773 if (digit >= 0 && digit <= 9) {
774 value = value * 10 + digit;
776 /* Now got 9 digits, so need to check
777 each time for overflow. */
779 while (digit >= 0 && digit <= 9
780 && (value < uv_max_div_10
781 || (value == uv_max_div_10
782 && digit <= uv_max_mod_10))) {
783 value = value * 10 + digit;
789 if (digit >= 0 && digit <= 9
792 skip the remaining digits, don't
793 worry about setting *valuep. */
796 } while (s < send && isDIGIT(*s));
798 IS_NUMBER_GREATER_THAN_UV_MAX;
818 numtype |= IS_NUMBER_IN_UV;
823 if (GROK_NUMERIC_RADIX(&s, send)) {
824 numtype |= IS_NUMBER_NOT_INT;
825 while (s < send && isDIGIT(*s)) /* optional digits after the radix */
829 else if (GROK_NUMERIC_RADIX(&s, send)) {
830 numtype |= IS_NUMBER_NOT_INT | IS_NUMBER_IN_UV; /* valuep assigned below */
831 /* no digits before the radix means we need digits after it */
832 if (s < send && isDIGIT(*s)) {
835 } while (s < send && isDIGIT(*s));
837 /* integer approximation is valid - it's 0. */
845 if (INFNAN_PEEK(d, send)) {
846 int infnan = Perl_grok_infnan(&d, send);
847 if ((infnan & IS_NUMBER_INFINITY)) {
851 else if ((infnan & IS_NUMBER_NAN)) {
862 /* Keep the sign for infinity. */
863 numtype |= IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT;
865 numtype &= IS_NUMBER_NEG; /* Clear sign for nan. */
866 numtype |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
867 } else if (s < send) {
868 /* we can have an optional exponent part */
869 if (isALPHA_FOLD_EQ(*s, 'e')) {
871 if (s < send && (*s == '-' || *s == '+'))
873 if (s < send && isDIGIT(*s)) {
876 } while (s < send && isDIGIT(*s));
878 else if (flags & PERL_SCAN_TRAILING)
879 return numtype | IS_NUMBER_TRAILING;
883 /* The only flag we keep is sign. Blow away any "it's UV" */
884 numtype &= IS_NUMBER_NEG;
885 numtype |= IS_NUMBER_NOT_INT;
888 while (s < send && isSPACE(*s))
892 if (len == 10 && memEQ(pv, "0 but true", 10)) {
895 return IS_NUMBER_IN_UV;
897 else if (flags & PERL_SCAN_TRAILING) {
898 return numtype | IS_NUMBER_TRAILING;
905 =for apidoc grok_atou
907 grok_atou is a safer replacement for atoi and strtol.
909 grok_atou parses a C-style zero-byte terminated string, looking for
910 a decimal unsigned integer.
912 Returns the unsigned integer, if a valid value can be parsed
913 from the beginning of the string.
915 Accepts only the decimal digits '0'..'9'.
917 As opposed to atoi or strtol, grok_atou does NOT allow optional
918 leading whitespace, or negative inputs. If such features are
919 required, the calling code needs to explicitly implement those.
921 If a valid value cannot be parsed, returns either zero (if non-digits
922 are met before any digits) or UV_MAX (if the value overflows).
924 Note that extraneous leading zeros also count as an overflow
925 (meaning that only "0" is the zero).
927 On failure, the *endptr is also set to NULL, unless endptr is NULL.
929 Trailing non-digit bytes are allowed if the endptr is non-NULL.
930 On return the *endptr will contain the pointer to the first non-digit byte.
932 If the endptr is NULL, the first non-digit byte MUST be
933 the zero byte terminating the pv, or zero will be returned.
935 Background: atoi has severe problems with illegal inputs, it cannot be
936 used for incremental parsing, and therefore should be avoided
937 atoi and strtol are also affected by locale settings, which can also be
938 seen as a bug (global state controlled by user environment).
944 Perl_grok_atou(const char *pv, const char** endptr)
948 const char* end2; /* Used in case endptr is NULL. */
949 UV val = 0; /* The return value. */
951 PERL_ARGS_ASSERT_GROK_ATOU;
953 eptr = endptr ? endptr : &end2;
955 /* Single-digit inputs are quite common. */
958 /* Extra leading zeros cause overflow. */
963 while (isDIGIT(*s)) {
964 /* This could be unrolled like in grok_number(), but
965 * the expected uses of this are not speed-needy, and
966 * unlikely to need full 64-bitness. */
967 U8 digit = *s++ - '0';
968 if (val < uv_max_div_10 ||
969 (val == uv_max_div_10 && digit <= uv_max_mod_10)) {
970 val = val * 10 + digit;
979 *eptr = NULL; /* If no progress, failed to parse anything. */
982 if (endptr == NULL && *s) {
983 return 0; /* If endptr is NULL, no trailing non-digits allowed. */
990 S_mulexp10(NV value, I32 exponent)
1002 /* On OpenVMS VAX we by default use the D_FLOAT double format,
1003 * and that format does not have *easy* capabilities [1] for
1004 * overflowing doubles 'silently' as IEEE fp does. We also need
1005 * to support G_FLOAT on both VAX and Alpha, and though the exponent
1006 * range is much larger than D_FLOAT it still doesn't do silent
1007 * overflow. Therefore we need to detect early whether we would
1008 * overflow (this is the behaviour of the native string-to-float
1009 * conversion routines, and therefore of native applications, too).
1011 * [1] Trying to establish a condition handler to trap floating point
1012 * exceptions is not a good idea. */
1014 /* In UNICOS and in certain Cray models (such as T90) there is no
1015 * IEEE fp, and no way at all from C to catch fp overflows gracefully.
1016 * There is something you can do if you are willing to use some
1017 * inline assembler: the instruction is called DFI-- but that will
1018 * disable *all* floating point interrupts, a little bit too large
1019 * a hammer. Therefore we need to catch potential overflows before
1022 #if ((defined(VMS) && !defined(_IEEE_FP)) || defined(_UNICOS)) && defined(NV_MAX_10_EXP)
1024 const NV exp_v = log10(value);
1025 if (exponent >= NV_MAX_10_EXP || exponent + exp_v >= NV_MAX_10_EXP)
1028 if (-(exponent + exp_v) >= NV_MAX_10_EXP)
1030 while (-exponent >= NV_MAX_10_EXP) {
1031 /* combination does not overflow, but 10^(-exponent) does */
1041 exponent = -exponent;
1042 #ifdef NV_MAX_10_EXP
1043 /* for something like 1234 x 10^-309, the action of calculating
1044 * the intermediate value 10^309 then returning 1234 / (10^309)
1045 * will fail, since 10^309 becomes infinity. In this case try to
1046 * refactor it as 123 / (10^308) etc.
1048 while (value && exponent > NV_MAX_10_EXP) {
1054 for (bit = 1; exponent; bit <<= 1) {
1055 if (exponent & bit) {
1058 /* Floating point exceptions are supposed to be turned off,
1059 * but if we're obviously done, don't risk another iteration.
1061 if (exponent == 0) break;
1065 return negative ? value / result : value * result;
1069 Perl_my_atof(pTHX_ const char* s)
1072 #ifdef USE_LOCALE_NUMERIC
1073 PERL_ARGS_ASSERT_MY_ATOF;
1076 DECLARE_STORE_LC_NUMERIC_SET_TO_NEEDED();
1077 if (PL_numeric_radix_sv && IN_LC(LC_NUMERIC)) {
1078 const char *standard = NULL, *local = NULL;
1079 bool use_standard_radix;
1081 /* Look through the string for the first thing that looks like a
1082 * decimal point: either the value in the current locale or the
1083 * standard fallback of '.'. The one which appears earliest in the
1084 * input string is the one that we should have atof look for. Note
1085 * that we have to determine this beforehand because on some
1086 * systems, Perl_atof2 is just a wrapper around the system's atof.
1088 standard = strchr(s, '.');
1089 local = strstr(s, SvPV_nolen(PL_numeric_radix_sv));
1091 use_standard_radix = standard && (!local || standard < local);
1093 if (use_standard_radix)
1094 SET_NUMERIC_STANDARD();
1098 if (use_standard_radix)
1099 SET_NUMERIC_LOCAL();
1103 RESTORE_LC_NUMERIC();
1112 Perl_my_atof2(pTHX_ const char* orig, NV* value)
1114 NV result[3] = {0.0, 0.0, 0.0};
1115 const char* s = orig;
1116 #ifdef USE_PERL_ATOF
1117 UV accumulator[2] = {0,0}; /* before/after dp */
1119 const char* send = s + strlen(orig); /* one past the last */
1120 bool seen_digit = 0;
1121 I32 exp_adjust[2] = {0,0};
1122 I32 exp_acc[2] = {-1, -1};
1123 /* the current exponent adjust for the accumulators */
1128 I32 sig_digits = 0; /* noof significant digits seen so far */
1130 PERL_ARGS_ASSERT_MY_ATOF2;
1132 /* There is no point in processing more significant digits
1133 * than the NV can hold. Note that NV_DIG is a lower-bound value,
1134 * while we need an upper-bound value. We add 2 to account for this;
1135 * since it will have been conservative on both the first and last digit.
1136 * For example a 32-bit mantissa with an exponent of 4 would have
1137 * exact values in the set
1145 * where for the purposes of calculating NV_DIG we would have to discount
1146 * both the first and last digit, since neither can hold all values from
1147 * 0..9; but for calculating the value we must examine those two digits.
1149 #ifdef MAX_SIG_DIG_PLUS
1150 /* It is not necessarily the case that adding 2 to NV_DIG gets all the
1151 possible digits in a NV, especially if NVs are not IEEE compliant
1152 (e.g., long doubles on IRIX) - Allen <allens@cpan.org> */
1153 # define MAX_SIG_DIGITS (NV_DIG+MAX_SIG_DIG_PLUS)
1155 # define MAX_SIG_DIGITS (NV_DIG+2)
1158 /* the max number we can accumulate in a UV, and still safely do 10*N+9 */
1159 #define MAX_ACCUMULATE ( (UV) ((UV_MAX - 9)/10))
1161 /* leading whitespace */
1175 const char *p0 = negative ? s - 1 : s;
1177 #if defined(NV_INF) && defined(NV_NAN)
1178 int infnan_flags = grok_infnan(&p, send);
1179 if (infnan_flags && p != p0) {
1180 if ((infnan_flags & IS_NUMBER_INFINITY)) {
1181 *value = (infnan_flags & IS_NUMBER_NEG) ? -NV_INF: NV_INF;
1184 else if ((infnan_flags & IS_NUMBER_NAN)) {
1189 #elif defined(HAS_STRTOD)
1190 if (INFNAN_PEEK(s, send)) {
1191 /* The native strtod() may not get all the possible
1192 * inf/nan strings INFNAN_PEEK() recognizes. */
1194 NV nv = Perl_strtod(p, &endp);
1203 /* we accumulate digits into an integer; when this becomes too
1204 * large, we add the total to NV and start again */
1214 /* don't start counting until we see the first significant
1215 * digit, eg the 5 in 0.00005... */
1216 if (!sig_digits && digit == 0)
1219 if (++sig_digits > MAX_SIG_DIGITS) {
1220 /* limits of precision reached */
1222 ++accumulator[seen_dp];
1223 } else if (digit == 5) {
1224 if (old_digit % 2) { /* round to even - Allen */
1225 ++accumulator[seen_dp];
1233 /* skip remaining digits */
1234 while (isDIGIT(*s)) {
1240 /* warn of loss of precision? */
1243 if (accumulator[seen_dp] > MAX_ACCUMULATE) {
1244 /* add accumulator to result and start again */
1245 result[seen_dp] = S_mulexp10(result[seen_dp],
1247 + (NV)accumulator[seen_dp];
1248 accumulator[seen_dp] = 0;
1249 exp_acc[seen_dp] = 0;
1251 accumulator[seen_dp] = accumulator[seen_dp] * 10 + digit;
1255 else if (!seen_dp && GROK_NUMERIC_RADIX(&s, send)) {
1257 if (sig_digits > MAX_SIG_DIGITS) {
1260 } while (isDIGIT(*s));
1269 result[0] = S_mulexp10(result[0], exp_acc[0]) + (NV)accumulator[0];
1271 result[1] = S_mulexp10(result[1], exp_acc[1]) + (NV)accumulator[1];
1274 if (seen_digit && (isALPHA_FOLD_EQ(*s, 'e'))) {
1275 bool expnegative = 0;
1286 exponent = exponent * 10 + (*s++ - '0');
1288 exponent = -exponent;
1293 /* now apply the exponent */
1296 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0])
1297 + S_mulexp10(result[1],exponent-exp_adjust[1]);
1299 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0]);
1302 /* now apply the sign */
1304 result[2] = -result[2];
1305 #endif /* USE_PERL_ATOF */
1310 #if ! defined(HAS_MODFL) && defined(HAS_AINTL) && defined(HAS_COPYSIGNL)
1312 Perl_my_modfl(long double x, long double *ip)
1315 return (x == *ip ? copysignl(0.0L, x) : x - *ip);
1319 #if ! defined(HAS_FREXPL) && defined(HAS_ILOGBL) && defined(HAS_SCALBNL)
1321 Perl_my_frexpl(long double x, int *e) {
1322 *e = x == 0.0L ? 0 : ilogbl(x) + 1;
1323 return (scalbnl(x, -*e));
1328 =for apidoc Perl_signbit
1330 Return a non-zero integer if the sign bit on an NV is set, and 0 if
1333 If Configure detects this system has a signbit() that will work with
1334 our NVs, then we just use it via the #define in perl.h. Otherwise,
1335 fall back on this implementation. The main use of this function
1338 Configure notes: This function is called 'Perl_signbit' instead of a
1339 plain 'signbit' because it is easy to imagine a system having a signbit()
1340 function or macro that doesn't happen to work with our particular choice
1341 of NVs. We shouldn't just re-#define signbit as Perl_signbit and expect
1342 the standard system headers to be happy. Also, this is a no-context
1343 function (no pTHX_) because Perl_signbit() is usually re-#defined in
1344 perl.h as a simple macro call to the system's signbit().
1345 Users should just always call Perl_signbit().
1349 #if !defined(HAS_SIGNBIT)
1351 Perl_signbit(NV x) {
1352 # ifdef Perl_fp_class_nzero
1354 return Perl_fp_class_nzero(x);
1356 return (x < 0.0) ? 1 : 0;
1362 * c-indentation-style: bsd
1364 * indent-tabs-mode: nil
1367 * ex: set ts=8 sts=4 sw=4 et: