sysdeps/ieee754/dbl-64/s_log1p.c

   1 /* @(#)s_log1p.c 5.1 93/09/24 */
   2 /*
   3  * ====================================================
   4  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
   5  *
   6  * Developed at SunPro, a Sun Microsystems, Inc. business.
   7  * Permission to use, copy, modify, and distribute this
   8  * software is freely granted, provided that this notice
   9  * is preserved.
  10  * ====================================================
  11  */
  12 /* Modified by Naohiko Shimizu/Tokai University, Japan 1997/08/25,
  13    for performance improvement on pipelined processors.
  14 */
  15
  16 #if defined(LIBM_SCCS) && !defined(lint)
  17 static char rcsid[] = "$NetBSD: s_log1p.c,v 1.8 1995/05/10 20:47:46 jtc Exp $";
  18 #endif
  19
  20 /* double log1p(double x)
  21  *
  22  * Method :
  23  *   1. Argument Reduction: find k and f such that
  24  *                      1+x = 2^k * (1+f),
  25  *         where  sqrt(2)/2 < 1+f < sqrt(2) .
  26  *
  27  *      Note. If k=0, then f=x is exact. However, if k!=0, then f
  28  *      may not be representable exactly. In that case, a correction
  29  *      term is need. Let u=1+x rounded. Let c = (1+x)-u, then
  30  *      log(1+x) - log(u) ~ c/u. Thus, we proceed to compute log(u),
  31  *      and add back the correction term c/u.
  32  *      (Note: when x > 2**53, one can simply return log(x))
  33  *
  34  *   2. Approximation of log1p(f).
  35  *      Let s = f/(2+f) ; based on log(1+f) = log(1+s) - log(1-s)
  36  *               = 2s + 2/3 s**3 + 2/5 s**5 + .....,
  37  *               = 2s + s*R
  38  *      We use a special Reme algorithm on [0,0.1716] to generate
  39  *      a polynomial of degree 14 to approximate R The maximum error
  40  *      of this polynomial approximation is bounded by 2**-58.45. In
  41  *      other words,
  42  *                      2      4      6      8      10      12      14
  43  *          R(z) ~ Lp1*s +Lp2*s +Lp3*s +Lp4*s +Lp5*s  +Lp6*s  +Lp7*s
  44  *      (the values of Lp1 to Lp7 are listed in the program)
  45  *      and
  46  *          |      2          14          |     -58.45
  47  *          | Lp1*s +...+Lp7*s    -  R(z) | <= 2
  48  *          |                             |
  49  *      Note that 2s = f - s*f = f - hfsq + s*hfsq, where hfsq = f*f/2.
  50  *      In order to guarantee error in log below 1ulp, we compute log
  51  *      by
  52  *              log1p(f) = f - (hfsq - s*(hfsq+R)).
  53  *
  54  *      3. Finally, log1p(x) = k*ln2 + log1p(f).
  55  *                           = k*ln2_hi+(f-(hfsq-(s*(hfsq+R)+k*ln2_lo)))
  56  *         Here ln2 is split into two floating point number:
  57  *                      ln2_hi + ln2_lo,
  58  *         where n*ln2_hi is always exact for |n| < 2000.
  59  *
  60  * Special cases:
  61  *      log1p(x) is NaN with signal if x < -1 (including -INF) ;
  62  *      log1p(+INF) is +INF; log1p(-1) is -INF with signal;
  63  *      log1p(NaN) is that NaN with no signal.
  64  *
  65  * Accuracy:
  66  *      according to an error analysis, the error is always less than
  67  *      1 ulp (unit in the last place).
  68  *
  69  * Constants:
  70  * The hexadecimal values are the intended ones for the following
  71  * constants. The decimal values may be used, provided that the
  72  * compiler will convert from decimal to binary accurately enough
  73  * to produce the hexadecimal values shown.
  74  *
  75  * Note: Assuming log() return accurate answer, the following
  76  *       algorithm can be used to compute log1p(x) to within a few ULP:
  77  *
  78  *              u = 1+x;
  79  *              if(u==1.0) return x ; else
  80  *                         return log(u)*(x/(u-1.0));
  81  *
  82  *       See HP-15C Advanced Functions Handbook, p.193.
  83  */
  84
  85 #include "math.h"
  86 #include "math_private.h"
  87
  88 #ifdef __STDC__
  89 static const double
  90 #else
  91 static double
  92 #endif
  93 ln2_hi  =  6.93147180369123816490e-01,  /* 3fe62e42 fee00000 */
  94 ln2_lo  =  1.90821492927058770002e-10,  /* 3dea39ef 35793c76 */
  95 two54   =  1.80143985094819840000e+16,  /* 43500000 00000000 */
  96 Lp[] = {0.0, 6.666666666666735130e-01,  /* 3FE55555 55555593 */
  97  3.999999999940941908e-01,  /* 3FD99999 9997FA04 */
  98  2.857142874366239149e-01,  /* 3FD24924 94229359 */
  99  2.222219843214978396e-01,  /* 3FCC71C5 1D8E78AF */
 100  1.818357216161805012e-01,  /* 3FC74664 96CB03DE */
 101  1.531383769920937332e-01,  /* 3FC39A09 D078C69F */
 102  1.479819860511658591e-01};  /* 3FC2F112 DF3E5244 */
 103
 104 #ifdef __STDC__
 105 static const double zero = 0.0;
 106 #else
 107 static double zero = 0.0;
 108 #endif
 109
 110 #ifdef __STDC__
 111         double __log1p(double x)
 112 #else
 113         double __log1p(x)
 114         double x;
 115 #endif
 116 {
 117         double hfsq,f,c,s,z,R,u,z2,z4,z6,R1,R2,R3,R4;
 118         int32_t k,hx,hu,ax;
 119
 120         GET_HIGH_WORD(hx,x);
 121         ax = hx&0x7fffffff;
 122
 123         k = 1;
 124         if (hx < 0x3FDA827A) {                  /* x < 0.41422  */
 125             if(ax>=0x3ff00000) {                /* x <= -1.0 */
 126                 if(x==-1.0) return -two54/(x-x);/* log1p(-1)=+inf */
 127                 else return (x-x)/(x-x);        /* log1p(x<-1)=NaN */
 128             }
 129             if(ax<0x3e200000) {                 /* |x| < 2**-29 */
 130                 if(two54+x>zero                 /* raise inexact */
 131                     &&ax<0x3c900000)            /* |x| < 2**-54 */
 132                     return x;
 133                 else
 134                     return x - x*x*0.5;
 135             }
 136             if(hx>0||hx<=((int32_t)0xbfd2bec3)) {
 137                 k=0;f=x;hu=1;}  /* -0.2929<x<0.41422 */
 138         }
 139         if (hx >= 0x7ff00000) return x+x;
 140         if(k!=0) {
 141             if(hx<0x43400000) {
 142                 u  = 1.0+x;
 143                 GET_HIGH_WORD(hu,u);
 144                 k  = (hu>>20)-1023;
 145                 c  = (k>0)? 1.0-(u-x):x-(u-1.0);/* correction term */
 146                 c /= u;
 147             } else {
 148                 u  = x;
 149                 GET_HIGH_WORD(hu,u);
 150                 k  = (hu>>20)-1023;
 151                 c  = 0;
 152             }
 153             hu &= 0x000fffff;
 154             if(hu<0x6a09e) {
 155                 SET_HIGH_WORD(u,hu|0x3ff00000); /* normalize u */
 156             } else {
 157                 k += 1;
 158                 SET_HIGH_WORD(u,hu|0x3fe00000); /* normalize u/2 */
 159                 hu = (0x00100000-hu)>>2;
 160             }
 161             f = u-1.0;
 162         }
 163         hfsq=0.5*f*f;
 164         if(hu==0) {     /* |f| < 2**-20 */
 165             if(f==zero) {
 166               if(k==0) return zero;
 167                         else {c += k*ln2_lo; return k*ln2_hi+c;}
 168             }
 169             R = hfsq*(1.0-0.66666666666666666*f);
 170             if(k==0) return f-R; else
 171                      return k*ln2_hi-((R-(k*ln2_lo+c))-f);
 172         }
 173         s = f/(2.0+f);
 174         z = s*s;
 175 #ifdef DO_NOT_USE_THIS
 176         R = z*(Lp1+z*(Lp2+z*(Lp3+z*(Lp4+z*(Lp5+z*(Lp6+z*Lp7))))));
 177 #else
 178         R1 = z*Lp[1]; z2=z*z;
 179         R2 = Lp[2]+z*Lp[3]; z4=z2*z2;
 180         R3 = Lp[4]+z*Lp[5]; z6=z4*z2;
 181         R4 = Lp[6]+z*Lp[7];
 182         R = R1 + z2*R2 + z4*R3 + z6*R4;
 183 #endif
 184         if(k==0) return f-(hfsq-s*(hfsq+R)); else
 185                  return k*ln2_hi-((hfsq-(s*(hfsq+R)+(k*ln2_lo+c)))-f);
 186 }
 187 weak_alias (__log1p, log1p)
 188 #ifdef NO_LONG_DOUBLE
 189 strong_alias (__log1p, __log1pl)
 190 weak_alias (__log1p, log1pl)
 191 #endif