3 * Copyright (c) 2001-2002, Larry Wall
5 * You may distribute under the terms of either the GNU General Public
6 * License or the Artistic License, as specified in the README file.
11 * "That only makes eleven (plus one mislaid) and not fourteen, unless
12 * wizards count differently to other people."
16 =head1 Numeric functions
20 #define PERL_IN_NUMERIC_C
24 Perl_cast_ulong(pTHX_ NV f)
27 return f < I32_MIN ? (U32) I32_MIN : (U32)(I32) f;
30 if (f < U32_MAX_P1_HALF)
33 return ((U32) f) | (1 + U32_MAX >> 1);
38 return f > 0 ? U32_MAX : 0 /* NaN */;
42 Perl_cast_i32(pTHX_ NV f)
45 return f < I32_MIN ? I32_MIN : (I32) f;
48 if (f < U32_MAX_P1_HALF)
51 return (I32)(((U32) f) | (1 + U32_MAX >> 1));
56 return f > 0 ? (I32)U32_MAX : 0 /* NaN */;
60 Perl_cast_iv(pTHX_ NV f)
63 return f < IV_MIN ? IV_MIN : (IV) f;
66 /* For future flexibility allowing for sizeof(UV) >= sizeof(IV) */
67 if (f < UV_MAX_P1_HALF)
70 return (IV)(((UV) f) | (1 + UV_MAX >> 1));
75 return f > 0 ? (IV)UV_MAX : 0 /* NaN */;
79 Perl_cast_uv(pTHX_ NV f)
82 return f < IV_MIN ? (UV) IV_MIN : (UV)(IV) f;
85 if (f < UV_MAX_P1_HALF)
88 return ((UV) f) | (1 + UV_MAX >> 1);
93 return f > 0 ? UV_MAX : 0 /* NaN */;
96 #if defined(HUGE_VAL) || (defined(USE_LONG_DOUBLE) && defined(HUGE_VALL))
98 * This hack is to force load of "huge" support from libm.a
99 * So it is in perl for (say) POSIX to use.
100 * Needed for SunOS with Sun's 'acc' for example.
105 # if defined(USE_LONG_DOUBLE) && defined(HUGE_VALL)
115 converts a string representing a binary number to numeric form.
117 On entry I<start> and I<*len> give the string to scan, I<*flags> gives
118 conversion flags, and I<result> should be NULL or a pointer to an NV.
119 The scan stops at the end of the string, or the first invalid character.
120 On return I<*len> is set to the length scanned string, and I<*flags> gives
123 If the value is <= UV_MAX it is returned as a UV, the output flags are clear,
124 and nothing is written to I<*result>. If the value is > UV_MAX C<grok_bin>
125 returns UV_MAX, sets C<PERL_SCAN_GREATER_THAN_UV_MAX> in the output flags,
126 and writes the value to I<*result> (or the value is discarded if I<result>
129 The hex number may optionally be prefixed with "0b" or "b" unless
130 C<PERL_SCAN_DISALLOW_PREFIX> is set in I<*flags> on entry. If
131 C<PERL_SCAN_ALLOW_UNDERSCORES> is set in I<*flags> then the binary
132 number may use '_' characters to separate digits.
138 Perl_grok_bin(pTHX_ char *start, STRLEN *len_p, I32 *flags, NV *result) {
139 const char *s = start;
144 const UV max_div_2 = UV_MAX / 2;
145 bool allow_underscores = *flags & PERL_SCAN_ALLOW_UNDERSCORES;
146 bool overflowed = FALSE;
148 if (!(*flags & PERL_SCAN_DISALLOW_PREFIX)) {
149 /* strip off leading b or 0b.
150 for compatibility silently suffer "b" and "0b" as valid binary
157 else if (len >= 2 && s[0] == '0' && s[1] == 'b') {
164 for (; len-- && *s; s++) {
166 if (bit == '0' || bit == '1') {
167 /* Write it in this wonky order with a goto to attempt to get the
168 compiler to make the common case integer-only loop pretty tight.
169 With gcc seems to be much straighter code than old scan_bin. */
172 if (value <= max_div_2) {
173 value = (value << 1) | (bit - '0');
176 /* Bah. We're just overflowed. */
177 if (ckWARN_d(WARN_OVERFLOW))
178 Perl_warner(aTHX_ packWARN(WARN_OVERFLOW),
179 "Integer overflow in binary number");
181 value_nv = (NV) value;
184 /* If an NV has not enough bits in its mantissa to
185 * represent a UV this summing of small low-order numbers
186 * is a waste of time (because the NV cannot preserve
187 * the low-order bits anyway): we could just remember when
188 * did we overflow and in the end just multiply value_nv by the
190 value_nv += (NV)(bit - '0');
193 if (bit == '_' && len && allow_underscores && (bit = s[1])
194 && (bit == '0' || bit == '1'))
200 if (!(*flags & PERL_SCAN_SILENT_ILLDIGIT) && ckWARN(WARN_DIGIT))
201 Perl_warner(aTHX_ packWARN(WARN_DIGIT),
202 "Illegal binary digit '%c' ignored", *s);
206 if ( ( overflowed && value_nv > 4294967295.0)
208 || (!overflowed && value > 0xffffffff )
211 if (ckWARN(WARN_PORTABLE))
212 Perl_warner(aTHX_ packWARN(WARN_PORTABLE),
213 "Binary number > 0b11111111111111111111111111111111 non-portable");
220 *flags = PERL_SCAN_GREATER_THAN_UV_MAX;
229 converts a string representing a hex number to numeric form.
231 On entry I<start> and I<*len> give the string to scan, I<*flags> gives
232 conversion flags, and I<result> should be NULL or a pointer to an NV.
233 The scan stops at the end of the string, or the first non-hex-digit character.
234 On return I<*len> is set to the length scanned string, and I<*flags> gives
237 If the value is <= UV_MAX it is returned as a UV, the output flags are clear,
238 and nothing is written to I<*result>. If the value is > UV_MAX C<grok_hex>
239 returns UV_MAX, sets C<PERL_SCAN_GREATER_THAN_UV_MAX> in the output flags,
240 and writes the value to I<*result> (or the value is discarded if I<result>
243 The hex number may optionally be prefixed with "0x" or "x" unless
244 C<PERL_SCAN_DISALLOW_PREFIX> is set in I<*flags> on entry. If
245 C<PERL_SCAN_ALLOW_UNDERSCORES> is set in I<*flags> then the hex
246 number may use '_' characters to separate digits.
252 Perl_grok_hex(pTHX_ char *start, STRLEN *len_p, I32 *flags, NV *result) {
253 const char *s = start;
258 const UV max_div_16 = UV_MAX / 16;
259 bool allow_underscores = *flags & PERL_SCAN_ALLOW_UNDERSCORES;
260 bool overflowed = FALSE;
261 const char *hexdigit;
263 if (!(*flags & PERL_SCAN_DISALLOW_PREFIX)) {
264 /* strip off leading x or 0x.
265 for compatibility silently suffer "x" and "0x" as valid hex numbers.
272 else if (len >= 2 && s[0] == '0' && s[1] == 'x') {
279 for (; len-- && *s; s++) {
280 hexdigit = strchr((char *) PL_hexdigit, *s);
282 /* Write it in this wonky order with a goto to attempt to get the
283 compiler to make the common case integer-only loop pretty tight.
284 With gcc seems to be much straighter code than old scan_hex. */
287 if (value <= max_div_16) {
288 value = (value << 4) | ((hexdigit - PL_hexdigit) & 15);
291 /* Bah. We're just overflowed. */
292 if (ckWARN_d(WARN_OVERFLOW))
293 Perl_warner(aTHX_ packWARN(WARN_OVERFLOW),
294 "Integer overflow in hexadecimal number");
296 value_nv = (NV) value;
299 /* If an NV has not enough bits in its mantissa to
300 * represent a UV this summing of small low-order numbers
301 * is a waste of time (because the NV cannot preserve
302 * the low-order bits anyway): we could just remember when
303 * did we overflow and in the end just multiply value_nv by the
304 * right amount of 16-tuples. */
305 value_nv += (NV)((hexdigit - PL_hexdigit) & 15);
308 if (*s == '_' && len && allow_underscores && s[1]
309 && (hexdigit = strchr((char *) PL_hexdigit, s[1])))
315 if (!(*flags & PERL_SCAN_SILENT_ILLDIGIT) && ckWARN(WARN_DIGIT))
316 Perl_warner(aTHX_ packWARN(WARN_DIGIT),
317 "Illegal hexadecimal digit '%c' ignored", *s);
321 if ( ( overflowed && value_nv > 4294967295.0)
323 || (!overflowed && value > 0xffffffff )
326 if (ckWARN(WARN_PORTABLE))
327 Perl_warner(aTHX_ packWARN(WARN_PORTABLE),
328 "Hexadecimal number > 0xffffffff non-portable");
335 *flags = PERL_SCAN_GREATER_THAN_UV_MAX;
349 Perl_grok_oct(pTHX_ char *start, STRLEN *len_p, I32 *flags, NV *result) {
350 const char *s = start;
355 const UV max_div_8 = UV_MAX / 8;
356 bool allow_underscores = *flags & PERL_SCAN_ALLOW_UNDERSCORES;
357 bool overflowed = FALSE;
359 for (; len-- && *s; s++) {
360 /* gcc 2.95 optimiser not smart enough to figure that this subtraction
361 out front allows slicker code. */
362 int digit = *s - '0';
363 if (digit >= 0 && digit <= 7) {
364 /* Write it in this wonky order with a goto to attempt to get the
365 compiler to make the common case integer-only loop pretty tight.
369 if (value <= max_div_8) {
370 value = (value << 3) | digit;
373 /* Bah. We're just overflowed. */
374 if (ckWARN_d(WARN_OVERFLOW))
375 Perl_warner(aTHX_ packWARN(WARN_OVERFLOW),
376 "Integer overflow in octal number");
378 value_nv = (NV) value;
381 /* If an NV has not enough bits in its mantissa to
382 * represent a UV this summing of small low-order numbers
383 * is a waste of time (because the NV cannot preserve
384 * the low-order bits anyway): we could just remember when
385 * did we overflow and in the end just multiply value_nv by the
386 * right amount of 8-tuples. */
387 value_nv += (NV)digit;
390 if (digit == ('_' - '0') && len && allow_underscores
391 && (digit = s[1] - '0') && (digit >= 0 && digit <= 7))
397 /* Allow \octal to work the DWIM way (that is, stop scanning
398 * as soon as non-octal characters are seen, complain only iff
399 * someone seems to want to use the digits eight and nine). */
400 if (digit == 8 || digit == 9) {
401 if (!(*flags & PERL_SCAN_SILENT_ILLDIGIT) && ckWARN(WARN_DIGIT))
402 Perl_warner(aTHX_ packWARN(WARN_DIGIT),
403 "Illegal octal digit '%c' ignored", *s);
408 if ( ( overflowed && value_nv > 4294967295.0)
410 || (!overflowed && value > 0xffffffff )
413 if (ckWARN(WARN_PORTABLE))
414 Perl_warner(aTHX_ packWARN(WARN_PORTABLE),
415 "Octal number > 037777777777 non-portable");
422 *flags = PERL_SCAN_GREATER_THAN_UV_MAX;
431 For backwards compatibility. Use C<grok_bin> instead.
435 For backwards compatibility. Use C<grok_hex> instead.
439 For backwards compatibility. Use C<grok_oct> instead.
445 Perl_scan_bin(pTHX_ char *start, STRLEN len, STRLEN *retlen)
448 I32 flags = *retlen ? PERL_SCAN_ALLOW_UNDERSCORES : 0;
449 UV ruv = grok_bin (start, &len, &flags, &rnv);
452 return (flags & PERL_SCAN_GREATER_THAN_UV_MAX) ? rnv : (NV)ruv;
456 Perl_scan_oct(pTHX_ char *start, STRLEN len, STRLEN *retlen)
459 I32 flags = *retlen ? PERL_SCAN_ALLOW_UNDERSCORES : 0;
460 UV ruv = grok_oct (start, &len, &flags, &rnv);
463 return (flags & PERL_SCAN_GREATER_THAN_UV_MAX) ? rnv : (NV)ruv;
467 Perl_scan_hex(pTHX_ char *start, STRLEN len, STRLEN *retlen)
470 I32 flags = *retlen ? PERL_SCAN_ALLOW_UNDERSCORES : 0;
471 UV ruv = grok_hex (start, &len, &flags, &rnv);
474 return (flags & PERL_SCAN_GREATER_THAN_UV_MAX) ? rnv : (NV)ruv;
478 =for apidoc grok_numeric_radix
480 Scan and skip for a numeric decimal separator (radix).
485 Perl_grok_numeric_radix(pTHX_ const char **sp, const char *send)
487 #ifdef USE_LOCALE_NUMERIC
488 if (PL_numeric_radix_sv && IN_LOCALE) {
490 char* radix = SvPV(PL_numeric_radix_sv, len);
491 if (*sp + len <= send && memEQ(*sp, radix, len)) {
496 /* always try "." if numeric radix didn't match because
497 * we may have data from different locales mixed */
499 if (*sp < send && **sp == '.') {
507 =for apidoc grok_number
509 Recognise (or not) a number. The type of the number is returned
510 (0 if unrecognised), otherwise it is a bit-ORed combination of
511 IS_NUMBER_IN_UV, IS_NUMBER_GREATER_THAN_UV_MAX, IS_NUMBER_NOT_INT,
512 IS_NUMBER_NEG, IS_NUMBER_INFINITY, IS_NUMBER_NAN (defined in perl.h).
514 If the value of the number can fit an in UV, it is returned in the *valuep
515 IS_NUMBER_IN_UV will be set to indicate that *valuep is valid, IS_NUMBER_IN_UV
516 will never be set unless *valuep is valid, but *valuep may have been assigned
517 to during processing even though IS_NUMBER_IN_UV is not set on return.
518 If valuep is NULL, IS_NUMBER_IN_UV will be set for the same cases as when
519 valuep is non-NULL, but no actual assignment (or SEGV) will occur.
521 IS_NUMBER_NOT_INT will be set with IS_NUMBER_IN_UV if trailing decimals were
522 seen (in which case *valuep gives the true value truncated to an integer), and
523 IS_NUMBER_NEG if the number is negative (in which case *valuep holds the
524 absolute value). IS_NUMBER_IN_UV is not set if e notation was used or the
525 number is larger than a UV.
530 Perl_grok_number(pTHX_ const char *pv, STRLEN len, UV *valuep)
533 const char *send = pv + len;
534 const UV max_div_10 = UV_MAX / 10;
535 const char max_mod_10 = UV_MAX % 10;
540 while (s < send && isSPACE(*s))
544 } else if (*s == '-') {
546 numtype = IS_NUMBER_NEG;
554 /* next must be digit or the radix separator or beginning of infinity */
556 /* UVs are at least 32 bits, so the first 9 decimal digits cannot
559 /* This construction seems to be more optimiser friendly.
560 (without it gcc does the isDIGIT test and the *s - '0' separately)
561 With it gcc on arm is managing 6 instructions (6 cycles) per digit.
562 In theory the optimiser could deduce how far to unroll the loop
563 before checking for overflow. */
565 int digit = *s - '0';
566 if (digit >= 0 && digit <= 9) {
567 value = value * 10 + digit;
570 if (digit >= 0 && digit <= 9) {
571 value = value * 10 + digit;
574 if (digit >= 0 && digit <= 9) {
575 value = value * 10 + digit;
578 if (digit >= 0 && digit <= 9) {
579 value = value * 10 + digit;
582 if (digit >= 0 && digit <= 9) {
583 value = value * 10 + digit;
586 if (digit >= 0 && digit <= 9) {
587 value = value * 10 + digit;
590 if (digit >= 0 && digit <= 9) {
591 value = value * 10 + digit;
594 if (digit >= 0 && digit <= 9) {
595 value = value * 10 + digit;
597 /* Now got 9 digits, so need to check
598 each time for overflow. */
600 while (digit >= 0 && digit <= 9
601 && (value < max_div_10
602 || (value == max_div_10
603 && digit <= max_mod_10))) {
604 value = value * 10 + digit;
610 if (digit >= 0 && digit <= 9
613 skip the remaining digits, don't
614 worry about setting *valuep. */
617 } while (s < send && isDIGIT(*s));
619 IS_NUMBER_GREATER_THAN_UV_MAX;
639 numtype |= IS_NUMBER_IN_UV;
644 if (GROK_NUMERIC_RADIX(&s, send)) {
645 numtype |= IS_NUMBER_NOT_INT;
646 while (s < send && isDIGIT(*s)) /* optional digits after the radix */
650 else if (GROK_NUMERIC_RADIX(&s, send)) {
651 numtype |= IS_NUMBER_NOT_INT | IS_NUMBER_IN_UV; /* valuep assigned below */
652 /* no digits before the radix means we need digits after it */
653 if (s < send && isDIGIT(*s)) {
656 } while (s < send && isDIGIT(*s));
658 /* integer approximation is valid - it's 0. */
664 } else if (*s == 'I' || *s == 'i') {
665 s++; if (s == send || (*s != 'N' && *s != 'n')) return 0;
666 s++; if (s == send || (*s != 'F' && *s != 'f')) return 0;
667 s++; if (s < send && (*s == 'I' || *s == 'i')) {
668 s++; if (s == send || (*s != 'N' && *s != 'n')) return 0;
669 s++; if (s == send || (*s != 'I' && *s != 'i')) return 0;
670 s++; if (s == send || (*s != 'T' && *s != 't')) return 0;
671 s++; if (s == send || (*s != 'Y' && *s != 'y')) return 0;
675 } else if (*s == 'N' || *s == 'n') {
676 /* XXX TODO: There are signaling NaNs and quiet NaNs. */
677 s++; if (s == send || (*s != 'A' && *s != 'a')) return 0;
678 s++; if (s == send || (*s != 'N' && *s != 'n')) return 0;
685 numtype &= IS_NUMBER_NEG; /* Keep track of sign */
686 numtype |= IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT;
688 numtype &= IS_NUMBER_NEG; /* Keep track of sign */
689 numtype |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT;
690 } else if (s < send) {
691 /* we can have an optional exponent part */
692 if (*s == 'e' || *s == 'E') {
693 /* The only flag we keep is sign. Blow away any "it's UV" */
694 numtype &= IS_NUMBER_NEG;
695 numtype |= IS_NUMBER_NOT_INT;
697 if (s < send && (*s == '-' || *s == '+'))
699 if (s < send && isDIGIT(*s)) {
702 } while (s < send && isDIGIT(*s));
708 while (s < send && isSPACE(*s))
712 if (len == 10 && memEQ(pv, "0 but true", 10)) {
715 return IS_NUMBER_IN_UV;
721 S_mulexp10(NV value, I32 exponent)
733 /* On OpenVMS VAX we by default use the D_FLOAT double format,
734 * and that format does not have *easy* capabilities [1] for
735 * overflowing doubles 'silently' as IEEE fp does. We also need
736 * to support G_FLOAT on both VAX and Alpha, and though the exponent
737 * range is much larger than D_FLOAT it still doesn't do silent
738 * overflow. Therefore we need to detect early whether we would
739 * overflow (this is the behaviour of the native string-to-float
740 * conversion routines, and therefore of native applications, too).
742 * [1] Trying to establish a condition handler to trap floating point
743 * exceptions is not a good idea. */
745 /* In UNICOS and in certain Cray models (such as T90) there is no
746 * IEEE fp, and no way at all from C to catch fp overflows gracefully.
747 * There is something you can do if you are willing to use some
748 * inline assembler: the instruction is called DFI-- but that will
749 * disable *all* floating point interrupts, a little bit too large
750 * a hammer. Therefore we need to catch potential overflows before
753 #if ((defined(VMS) && !defined(__IEEE_FP)) || defined(_UNICOS)) && defined(NV_MAX_10_EXP)
755 NV exp_v = log10(value);
756 if (exponent >= NV_MAX_10_EXP || exponent + exp_v >= NV_MAX_10_EXP)
759 if (-(exponent + exp_v) >= NV_MAX_10_EXP)
761 while (-exponent >= NV_MAX_10_EXP) {
762 /* combination does not overflow, but 10^(-exponent) does */
772 exponent = -exponent;
774 for (bit = 1; exponent; bit <<= 1) {
775 if (exponent & bit) {
778 /* Floating point exceptions are supposed to be turned off,
779 * but if we're obviously done, don't risk another iteration.
781 if (exponent == 0) break;
785 return negative ? value / result : value * result;
789 Perl_my_atof(pTHX_ const char* s)
792 #ifdef USE_LOCALE_NUMERIC
793 if (PL_numeric_local && IN_LOCALE) {
796 /* Scan the number twice; once using locale and once without;
797 * choose the larger result (in absolute value). */
799 SET_NUMERIC_STANDARD();
802 if ((y < 0.0 && y < x) || (y > 0.0 && y > x))
814 Perl_my_atof2(pTHX_ const char* orig, NV* value)
816 NV result[3] = {0.0, 0.0, 0.0};
817 char* s = (char*)orig;
819 UV accumulator[2] = {0,0}; /* before/after dp */
821 char* send = s + strlen(orig) - 1;
823 I32 exp_adjust[2] = {0,0};
824 I32 exp_acc[2] = {-1, -1};
825 /* the current exponent adjust for the accumulators */
830 I32 sig_digits = 0; /* noof significant digits seen so far */
832 /* There is no point in processing more significant digits
833 * than the NV can hold. Note that NV_DIG is a lower-bound value,
834 * while we need an upper-bound value. We add 2 to account for this;
835 * since it will have been conservative on both the first and last digit.
836 * For example a 32-bit mantissa with an exponent of 4 would have
837 * exact values in the set
845 * where for the purposes of calculating NV_DIG we would have to discount
846 * both the first and last digit, since neither can hold all values from
847 * 0..9; but for calculating the value we must examine those two digits.
849 #define MAX_SIG_DIGITS (NV_DIG+2)
851 /* the max number we can accumulate in a UV, and still safely do 10*N+9 */
852 #define MAX_ACCUMULATE ( (UV) ((UV_MAX - 9)/10))
854 /* leading whitespace */
867 /* we accumulate digits into an integer; when this becomes too
868 * large, we add the total to NV and start again */
878 /* don't start counting until we see the first significant
879 * digit, eg the 5 in 0.00005... */
880 if (!sig_digits && digit == 0)
883 if (++sig_digits > MAX_SIG_DIGITS) {
884 /* limits of precision reached */
886 ++accumulator[seen_dp];
887 } else if (digit == 5) {
888 if (old_digit % 2) { /* round to even - Allen */
889 ++accumulator[seen_dp];
897 /* skip remaining digits */
898 while (isDIGIT(*s)) {
904 /* warn of loss of precision? */
907 if (accumulator[seen_dp] > MAX_ACCUMULATE) {
908 /* add accumulator to result and start again */
909 result[seen_dp] = S_mulexp10(result[seen_dp],
911 + (NV)accumulator[seen_dp];
912 accumulator[seen_dp] = 0;
913 exp_acc[seen_dp] = 0;
915 accumulator[seen_dp] = accumulator[seen_dp] * 10 + digit;
919 else if (!seen_dp && GROK_NUMERIC_RADIX((const char **)&s, send)) {
921 if (sig_digits > MAX_SIG_DIGITS) {
923 while (isDIGIT(*s)) {
934 result[0] = S_mulexp10(result[0], exp_acc[0]) + (NV)accumulator[0];
936 result[1] = S_mulexp10(result[1], exp_acc[1]) + (NV)accumulator[1];
939 if (seen_digit && (*s == 'e' || *s == 'E')) {
940 bool expnegative = 0;
951 exponent = exponent * 10 + (*s++ - '0');
953 exponent = -exponent;
958 /* now apply the exponent */
961 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0])
962 + S_mulexp10(result[1],exponent-exp_adjust[1]);
964 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0]);
967 /* now apply the sign */
969 result[2] = -result[2];
970 #endif /* USE_PERL_ATOF */