1 /* origin: FreeBSD /usr/src/lib/msun/src/s_expm1.c */
3 * ====================================================
4 * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
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
10 * ====================================================
13 * Returns exp(x)-1, the exponential of x minus 1.
16 * 1. Argument reduction:
17 * Given x, find r and integer k such that
19 * x = k*ln2 + r, |r| <= 0.5*ln2 ~ 0.34658
21 * Here a correction term c will be computed to compensate
22 * the error in r when rounded to a floating-point number.
24 * 2. Approximating expm1(r) by a special rational function on
25 * the interval [0,0.34658]:
27 * r*(exp(r)+1)/(exp(r)-1) = 2+ r^2/6 - r^4/360 + ...
28 * we define R1(r*r) by
29 * r*(exp(r)+1)/(exp(r)-1) = 2+ r^2/6 * R1(r*r)
31 * R1(r**2) = 6/r *((exp(r)+1)/(exp(r)-1) - 2/r)
32 * = 6/r * ( 1 + 2.0*(1/(exp(r)-1) - 1/r))
33 * = 1 - r^2/60 + r^4/2520 - r^6/100800 + ...
34 * We use a special Reme algorithm on [0,0.347] to generate
35 * a polynomial of degree 5 in r*r to approximate R1. The
36 * maximum error of this polynomial approximation is bounded
37 * by 2**-61. In other words,
38 * R1(z) ~ 1.0 + Q1*z + Q2*z**2 + Q3*z**3 + Q4*z**4 + Q5*z**5
39 * where Q1 = -1.6666666666666567384E-2,
40 * Q2 = 3.9682539681370365873E-4,
41 * Q3 = -9.9206344733435987357E-6,
42 * Q4 = 2.5051361420808517002E-7,
43 * Q5 = -6.2843505682382617102E-9;
45 * with error bounded by
47 * | 1.0+Q1*z+...+Q5*z - R1(z) | <= 2
50 * expm1(r) = exp(r)-1 is then computed by the following
51 * specific way which minimize the accumulation rounding error:
53 * r r [ 3 - (R1 + R1*r/2) ]
54 * expm1(r) = r + --- + --- * [--------------------]
55 * 2 2 [ 6 - r*(3 - R1*r/2) ]
57 * To compensate the error in the argument reduction, we use
58 * expm1(r+c) = expm1(r) + c + expm1(r)*c
59 * ~ expm1(r) + c + r*c
60 * Thus c+r*c will be added in as the correction terms for
61 * expm1(r+c). Now rearrange the term to avoid optimization
64 * ({ ( r [ R1 - (3 - R1*r/2) ] ) } r )
65 * expm1(r+c)~r - ({r*(--- * [--------------------]-c)-c} - --- )
66 * ({ ( 2 [ 6 - r*(3 - R1*r/2) ] ) } 2 )
70 * 3. Scale back to obtain expm1(x):
71 * From step 1, we have
72 * expm1(x) = either 2^k*[expm1(r)+1] - 1
73 * = or 2^k*[expm1(r) + (1-2^-k)]
74 * 4. Implementation notes:
75 * (A). To save one multiplication, we scale the coefficient Qi
76 * to Qi*2^i, and replace z by (x^2)/2.
77 * (B). To achieve maximum accuracy, we compute expm1(x) by
78 * (i) if x < -56*ln2, return -1.0, (raise inexact if x!=inf)
79 * (ii) if k=0, return r-E
80 * (iii) if k=-1, return 0.5*(r-E)-0.5
81 * (iv) if k=1 if r < -0.25, return 2*((r+0.5)- E)
82 * else return 1.0+2.0*(r-E);
83 * (v) if (k<-2||k>56) return 2^k(1-(E-r)) - 1 (or exp(x)-1)
84 * (vi) if k <= 20, return 2^k((1-2^-k)-(E-r)), else
85 * (vii) return 2^k(1-((E+2^-k)-r))
88 * expm1(INF) is INF, expm1(NaN) is NaN;
89 * expm1(-INF) is -1, and
90 * for finite argument, only expm1(0)=0 is exact.
93 * according to an error analysis, the error is always less than
94 * 1 ulp (unit in the last place).
98 * if x > 7.09782712893383973096e+02 then expm1(x) overflow
101 * The hexadecimal values are the intended ones for the following
102 * constants. The decimal values may be used, provided that the
103 * compiler will convert from decimal to binary accurately enough
104 * to produce the hexadecimal values shown.
112 o_threshold = 7.09782712893383973096e+02, /* 0x40862E42, 0xFEFA39EF */
113 ln2_hi = 6.93147180369123816490e-01, /* 0x3fe62e42, 0xfee00000 */
114 ln2_lo = 1.90821492927058770002e-10, /* 0x3dea39ef, 0x35793c76 */
115 invln2 = 1.44269504088896338700e+00, /* 0x3ff71547, 0x652b82fe */
116 /* Scaled Q's: Qn_here = 2**n * Qn_above, for R(2*z) where z = hxs = x*x/2: */
117 Q1 = -3.33333333333331316428e-02, /* BFA11111 111110F4 */
118 Q2 = 1.58730158725481460165e-03, /* 3F5A01A0 19FE5585 */
119 Q3 = -7.93650757867487942473e-05, /* BF14CE19 9EAADBB7 */
120 Q4 = 4.00821782732936239552e-06, /* 3ED0CFCA 86E65239 */
121 Q5 = -2.01099218183624371326e-07; /* BE8AFDB7 6E09C32D */
123 double expm1(double x)
125 double y,hi,lo,c,t,e,hxs,hfx,r1,twopk;
129 GET_HIGH_WORD(hx, x);
130 xsb = hx&0x80000000; /* sign bit of x */
131 hx &= 0x7fffffff; /* high word of |x| */
133 /* filter out huge and non-finite argument */
134 if (hx >= 0x4043687A) { /* if |x|>=56*ln2 */
135 if (hx >= 0x40862E42) { /* if |x|>=709.78... */
136 if (hx >= 0x7ff00000) {
139 GET_LOW_WORD(low, x);
140 if (((hx&0xfffff)|low) != 0) /* NaN */
142 return xsb==0 ? x : -1.0; /* exp(+-inf)={inf,-1} */
145 return huge*huge; /* overflow */
147 if (xsb != 0) { /* x < -56*ln2, return -1.0 with inexact */
150 return tiny-1.0; /* return -1 */
154 /* argument reduction */
155 if (hx > 0x3fd62e42) { /* if |x| > 0.5 ln2 */
156 if (hx < 0x3FF0A2B2) { /* and |x| < 1.5 ln2 */
167 k = invln2*x + (xsb==0 ? 0.5 : -0.5);
169 hi = x - t*ln2_hi; /* t*ln2_hi is exact here */
172 STRICT_ASSIGN(double, x, hi - lo);
174 } else if (hx < 0x3c900000) { /* |x| < 2**-54, return x */
175 /* raise inexact flags when x != 0 */
177 return x - (t-(huge+x));
181 /* x is now in primary range */
184 r1 = 1.0+hxs*(Q1+hxs*(Q2+hxs*(Q3+hxs*(Q4+hxs*Q5))));
186 e = hxs*((r1-t)/(6.0 - x*t));
187 if (k == 0) /* c is 0 */
188 return x - (x*e-hxs);
189 INSERT_WORDS(twopk, 0x3ff00000+(k<<20), 0); /* 2^k */
193 return 0.5*(x-e) - 0.5;
196 return -2.0*(e-(x+0.5));
197 return 1.0+2.0*(x-e);
199 if (k <= -2 || k > 56) { /* suffice to return exp(x)-1 */
209 SET_HIGH_WORD(t, 0x3ff00000 - (0x200000>>k)); /* t=1-2^-k */
213 SET_HIGH_WORD(t, ((0x3ff-k)<<20)); /* 2^-k */