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 /* double log1p(double x)
17  *
18  * Method :
19  *   1. Argument Reduction: find k and f such that
20  *			1+x = 2^k * (1+f),
21  *	   where  sqrt(2)/2 < 1+f < sqrt(2) .
22  *
23  *      Note. If k=0, then f=x is exact. However, if k!=0, then f
24  *	may not be representable exactly. In that case, a correction
25  *	term is need. Let u=1+x rounded. Let c = (1+x)-u, then
26  *	log(1+x) - log(u) ~ c/u. Thus, we proceed to compute log(u),
27  *	and add back the correction term c/u.
28  *	(Note: when x > 2**53, one can simply return log(x))
29  *
30  *   2. Approximation of log1p(f).
31  *	Let s = f/(2+f) ; based on log(1+f) = log(1+s) - log(1-s)
32  *		 = 2s + 2/3 s**3 + 2/5 s**5 + .....,
33  *		 = 2s + s*R
34  *      We use a special Reme algorithm on [0,0.1716] to generate
35  *	a polynomial of degree 14 to approximate R The maximum error
36  *	of this polynomial approximation is bounded by 2**-58.45. In
37  *	other words,
38  *			2      4      6      8      10      12      14
39  *	    R(z) ~ Lp1*s +Lp2*s +Lp3*s +Lp4*s +Lp5*s  +Lp6*s  +Lp7*s
40  *	(the values of Lp1 to Lp7 are listed in the program)
41  *	and
42  *	    |      2          14          |     -58.45
43  *	    | Lp1*s +...+Lp7*s    -  R(z) | <= 2
44  *	    |                             |
45  *	Note that 2s = f - s*f = f - hfsq + s*hfsq, where hfsq = f*f/2.
46  *	In order to guarantee error in log below 1ulp, we compute log
47  *	by
48  *		log1p(f) = f - (hfsq - s*(hfsq+R)).
49  *
50  *	3. Finally, log1p(x) = k*ln2 + log1p(f).
51  *			     = k*ln2_hi+(f-(hfsq-(s*(hfsq+R)+k*ln2_lo)))
52  *	   Here ln2 is split into two floating point number:
53  *			ln2_hi + ln2_lo,
54  *	   where n*ln2_hi is always exact for |n| < 2000.
55  *
56  * Special cases:
57  *	log1p(x) is NaN with signal if x < -1 (including -INF) ;
58  *	log1p(+INF) is +INF; log1p(-1) is -INF with signal;
59  *	log1p(NaN) is that NaN with no signal.
60  *
61  * Accuracy:
62  *	according to an error analysis, the error is always less than
63  *	1 ulp (unit in the last place).
64  *
65  * Constants:
66  * The hexadecimal values are the intended ones for the following
67  * constants. The decimal values may be used, provided that the
68  * compiler will convert from decimal to binary accurately enough
69  * to produce the hexadecimal values shown.
70  *
71  * Note: Assuming log() return accurate answer, the following
72  *	 algorithm can be used to compute log1p(x) to within a few ULP:
73  *
74  *		u = 1+x;
75  *		if(u==1.0) return x ; else
76  *			   return log(u)*(x/(u-1.0));
77  *
78  *	 See HP-15C Advanced Functions Handbook, p.193.
79  */
80 
81 #include <float.h>
82 #include <math.h>
83 #include <math-barriers.h>
84 #include <math_private.h>
85 #include <math-underflow.h>
86 #include <libc-diag.h>
87 
88 static const double
89   ln2_hi = 6.93147180369123816490e-01,  /* 3fe62e42 fee00000 */
90   ln2_lo = 1.90821492927058770002e-10,  /* 3dea39ef 35793c76 */
91   two54 = 1.80143985094819840000e+16,   /* 43500000 00000000 */
92   Lp[] = { 0.0, 6.666666666666735130e-01, /* 3FE55555 55555593 */
93 	   3.999999999940941908e-01, /* 3FD99999 9997FA04 */
94 	   2.857142874366239149e-01, /* 3FD24924 94229359 */
95 	   2.222219843214978396e-01, /* 3FCC71C5 1D8E78AF */
96 	   1.818357216161805012e-01, /* 3FC74664 96CB03DE */
97 	   1.531383769920937332e-01, /* 3FC39A09 D078C69F */
98 	   1.479819860511658591e-01 }; /* 3FC2F112 DF3E5244 */
99 
100 static const double zero = 0.0;
101 
102 double
__log1p(double x)103 __log1p (double x)
104 {
105   double hfsq, f, c, s, z, R, u, z2, z4, z6, R1, R2, R3, R4;
106   int32_t k, hx, hu, ax;
107 
108   GET_HIGH_WORD (hx, x);
109   ax = hx & 0x7fffffff;
110 
111   k = 1;
112   if (hx < 0x3FDA827A)                          /* x < 0.41422  */
113     {
114       if (__glibc_unlikely (ax >= 0x3ff00000))           /* x <= -1.0 */
115 	{
116 	  if (x == -1.0)
117 	    return -two54 / zero;               /* log1p(-1)=-inf */
118 	  else
119 	    return (x - x) / (x - x);           /* log1p(x<-1)=NaN */
120 	}
121       if (__glibc_unlikely (ax < 0x3e200000))           /* |x| < 2**-29 */
122 	{
123 	  math_force_eval (two54 + x);          /* raise inexact */
124 	  if (ax < 0x3c900000)                  /* |x| < 2**-54 */
125 	    {
126 	      math_check_force_underflow (x);
127 	      return x;
128 	    }
129 	  else
130 	    return x - x * x * 0.5;
131 	}
132       if (hx > 0 || hx <= ((int32_t) 0xbfd2bec3))
133 	{
134 	  k = 0; f = x; hu = 1;
135 	}                       /* -0.2929<x<0.41422 */
136     }
137   else if (__glibc_unlikely (hx >= 0x7ff00000))
138     return x + x;
139   if (k != 0)
140     {
141       if (hx < 0x43400000)
142 	{
143 	  u = 1.0 + x;
144 	  GET_HIGH_WORD (hu, u);
145 	  k = (hu >> 20) - 1023;
146 	  c = (k > 0) ? 1.0 - (u - x) : x - (u - 1.0); /* correction term */
147 	  c /= u;
148 	}
149       else
150 	{
151 	  u = x;
152 	  GET_HIGH_WORD (hu, u);
153 	  k = (hu >> 20) - 1023;
154 	  c = 0;
155 	}
156       hu &= 0x000fffff;
157       if (hu < 0x6a09e)
158 	{
159 	  SET_HIGH_WORD (u, hu | 0x3ff00000);   /* normalize u */
160 	}
161       else
162 	{
163 	  k += 1;
164 	  SET_HIGH_WORD (u, hu | 0x3fe00000);   /* normalize u/2 */
165 	  hu = (0x00100000 - hu) >> 2;
166 	}
167       f = u - 1.0;
168     }
169   hfsq = 0.5 * f * f;
170   if (hu == 0)          /* |f| < 2**-20 */
171     {
172       if (f == zero)
173 	{
174 	  if (k == 0)
175 	    return zero;
176 	  else
177 	    {
178 	      c += k * ln2_lo; return k * ln2_hi + c;
179 	    }
180 	}
181       R = hfsq * (1.0 - 0.66666666666666666 * f);
182       if (k == 0)
183 	return f - R;
184       else
185 	return k * ln2_hi - ((R - (k * ln2_lo + c)) - f);
186     }
187   s = f / (2.0 + f);
188   z = s * s;
189   R1 = z * Lp[1]; z2 = z * z;
190   R2 = Lp[2] + z * Lp[3]; z4 = z2 * z2;
191   R3 = Lp[4] + z * Lp[5]; z6 = z4 * z2;
192   R4 = Lp[6] + z * Lp[7];
193   R = R1 + z2 * R2 + z4 * R3 + z6 * R4;
194   if (k == 0)
195     return f - (hfsq - s * (hfsq + R));
196   else
197     {
198       /* With GCC 7 when compiling with -Os the compiler warns that c
199 	 might be used uninitialized.  This can't be true because k
200 	 must be 0 for c to be uninitialized and we handled that
201 	 computation earlier without using c.  */
202       DIAG_PUSH_NEEDS_COMMENT;
203       DIAG_IGNORE_Os_NEEDS_COMMENT (7, "-Wmaybe-uninitialized");
204       return k * ln2_hi - ((hfsq - (s * (hfsq + R) + (k * ln2_lo + c))) - f);
205       DIAG_POP_NEEDS_COMMENT;
206     }
207 }
208