Python/strtod.c

   1 #include "pyconfig.h"
   2
   3 /* comp.sources.misc strtod(), as posted in comp.lang.tcl,
   4    with bugfix for "123000.0" and acceptance of space after 'e' sign nuked.
   5
   6    ************************************************************
   7    * YOU MUST EDIT THE MACHINE-DEPENDENT DEFINITIONS BELOW!!! *
   8    ************************************************************
   9 */
  10
  11 /*  File   : stdtod.c (Modified version of str2dbl.c)
  12     Author : Richard A. O'Keefe @ Quintus Computer Systems, Inc.
  13     Updated: Tuesday August 2nd, 1988
  14     Defines: double strtod (char *str, char**ptr)
  15 */
  16
  17 /*  This is an implementation of the strtod() function described in the
  18     System V manuals, with a different name to avoid linker problems.
  19     All that str2dbl() does itself is check that the argument is well-formed
  20     and is in range.  It leaves the work of conversion to atof(), which is
  21     assumed to exist and deliver correct results (if they can be represented).
  22
  23     There are two reasons why this should be provided to the net:
  24     (a) some UNIX systems do not yet have strtod(), or do not have it
  25         available in the BSD "universe" (but they do have atof()).
  26     (b) some of the UNIX systems that *do* have it get it wrong.
  27         (some crash with large arguments, some assign the wrong *ptr value).
  28     There is a reason why *we* are providing it: we need a correct version
  29     of strtod(), and if we give this one away maybe someone will look for
  30     mistakes in it and fix them for us (:-).
  31 */
  32
  33 /*  The following constants are machine-specific.  MD{MIN,MAX}EXPT are
  34     integers and MD{MIN,MAX}FRAC are strings such that
  35         0.${MDMAXFRAC}e${MDMAXEXPT} is the largest representable double,
  36         0.${MDMINFRAC}e${MDMINEXPT} is the smallest representable +ve double
  37     MD{MIN,MAX}FRAC must not have any trailing zeros.
  38     The values here are for IEEE-754 64-bit floats.
  39     It is not perfectly clear to me whether an IEEE infinity should be
  40     returned for overflow, nor what a portable way of writing one is,
  41     so HUGE is just 0.MAXFRAC*10**MAXEXPT (this seems still to be the
  42     UNIX convention).
  43
  44     I do know about <values.h>, but the whole point of this file is that
  45     we can't always trust that stuff to be there or to be correct.
  46 */
  47 static  int     MDMINEXPT       = -323;
  48 static  char    MDMINFRAC[]     = "494065645841246544";
  49 static  double  ZERO            = 0.0;
  50
  51 static  int     MDMAXEXPT       = 309;
  52 static  char    MDMAXFRAC[]     = "17976931348623157";
  53 static  double  HUGE            = 1.7976931348623157e308;
  54
  55 extern  double  atof(const char *);             /* Only called when result known to be ok */
  56
  57 #ifdef HAVE_ERRNO_H
  58 #include <errno.h>
  59 #endif
  60 extern  int     errno;
  61
  62 double strtod(char *str, char **ptr)
  63 {
  64         int sign, scale, dotseen;
  65         int esign, expt;
  66         char *save;
  67         register char *sp, *dp;
  68         register int c;
  69         char *buforg, *buflim;
  70         char buffer[64];                /* 45-digit significant + */
  71                                         /* 13-digit exponent */
  72         sp = str;
  73         while (*sp == ' ') sp++;
  74         sign = 1;
  75         if (*sp == '-') sign -= 2, sp++;
  76         dotseen = 0, scale = 0;
  77         dp = buffer;
  78         *dp++ = '0'; *dp++ = '.';
  79         buforg = dp, buflim = buffer+48;
  80         for (save = sp; c = *sp; sp++)
  81             if (c == '.') {
  82                 if (dotseen) break;
  83                 dotseen++;
  84             } else
  85             if ((unsigned)(c-'0') > (unsigned)('9'-'0')) {
  86                 break;
  87             } else
  88             if (c == '0') {
  89                 if (dp != buforg) {
  90                     /* This is not the first digit, so we want to keep it */
  91                     if (dp < buflim) *dp++ = c;
  92                     if (!dotseen) scale++;
  93                 } else {
  94                     /* No non-zero digits seen yet */
  95                     /* If a . has been seen, scale must be adjusted */
  96                     if (dotseen) scale--;
  97                 }
  98             } else {
  99                 /* This is a nonzero digit, so we want to keep it */
 100                 if (dp < buflim) *dp++ = c;
 101                 /* If it precedes a ., scale must be adjusted */
 102                 if (!dotseen) scale++;
 103             }
 104         if (sp == save) {
 105             if (ptr) *ptr = str;
 106             errno = EDOM;               /* what should this be? */
 107             return ZERO;
 108         }
 109
 110         while (dp > buforg && dp[-1] == '0') --dp;
 111         if (dp == buforg) *dp++ = '0';
 112         *dp = '\0';
 113         /*  Now the contents of buffer are
 114             +--+--------+-+--------+
 115             |0.|fraction|\|leftover|
 116             +--+--------+-+--------+
 117                          ^dp points here
 118             where fraction begins with 0 iff it is "0", and has at most
 119             45 digits in it, and leftover is at least 16 characters.
 120         */
 121         save = sp, expt = 0, esign = 1;
 122         do {
 123             c = *sp++;
 124             if (c != 'e' && c != 'E') break;
 125             c = *sp++;
 126             if (c == '-') esign -= 2, c = *sp++; else
 127             if (c == '+' /* || c == ' ' */ ) c = *sp++;
 128             if ((unsigned)(c-'0') > (unsigned)('9'-'0')) break;
 129             while (c == '0') c = *sp++;
 130             for (; (unsigned)(c-'0') <= (unsigned)('9'-'0'); c = *sp++)
 131                 expt = expt*10 + c-'0';
 132             if (esign < 0) expt = -expt;
 133             save = sp-1;
 134         } while (0);
 135         if (ptr) *ptr = save;
 136         expt += scale;
 137         /*  Now the number is sign*0.fraction*10**expt  */
 138         errno = ERANGE;
 139         if (expt > MDMAXEXPT) {
 140             return HUGE*sign;
 141         } else
 142         if (expt == MDMAXEXPT) {
 143             if (strcmp(buforg, MDMAXFRAC) > 0) return HUGE*sign;
 144         } else
 145         if (expt < MDMINEXPT) {
 146             return ZERO*sign;
 147         } else
 148         if (expt == MDMINEXPT) {
 149             if (strcmp(buforg, MDMINFRAC) < 0) return ZERO*sign;
 150         }
 151         /*  We have now established that the number can be  */
 152         /*  represented without overflow or underflow  */
 153         (void) sprintf(dp, "E%d", expt);
 154         errno = 0;
 155         return atof(buffer)*sign;
 156 }