1 /* @(#)e_pow.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 13 #ifndef lint 14 static char rcsid[] = "$Id: e_pow.c,v 1.3 1994/02/18 02:25:39 jtc Exp $"; 15 #endif 16 17 /* __ieee754_pow(x,y) return x**y 18 * 19 * n 20 * Method: Let x = 2 * (1+f) 21 * 1. Compute and return log2(x) in two pieces: 22 * log2(x) = w1 + w2, 23 * where w1 has 53-24 = 29 bit trailing zeros. 24 * 2. Perform y*log2(x) = n+y' by simulating muti-precision 25 * arithmetic, where |y'|<=0.5. 26 * 3. Return x**y = 2**n*exp(y'*log2) 27 * 28 * Special cases: 29 * 1. (anything) ** 0 is 1 30 * 2. (anything) ** 1 is itself 31 * 3. (anything) ** NAN is NAN 32 * 4. NAN ** (anything except 0) is NAN 33 * 5. +-(|x| > 1) ** +INF is +INF 34 * 6. +-(|x| > 1) ** -INF is +0 35 * 7. +-(|x| < 1) ** +INF is +0 36 * 8. +-(|x| < 1) ** -INF is +INF 37 * 9. +-1 ** +-INF is NAN 38 * 10. +0 ** (+anything except 0, NAN) is +0 39 * 11. -0 ** (+anything except 0, NAN, odd integer) is +0 40 * 12. +0 ** (-anything except 0, NAN) is +INF 41 * 13. -0 ** (-anything except 0, NAN, odd integer) is +INF 42 * 14. -0 ** (odd integer) = -( +0 ** (odd integer) ) 43 * 15. +INF ** (+anything except 0,NAN) is +INF 44 * 16. +INF ** (-anything except 0,NAN) is +0 45 * 17. -INF ** (anything) = -0 ** (-anything) 46 * 18. (-anything) ** (integer) is (-1)**(integer)*(+anything**integer) 47 * 19. (-anything except 0 and inf) ** (non-integer) is NAN 48 * 49 * Accuracy: 50 * pow(x,y) returns x**y nearly rounded. In particular 51 * pow(integer,integer) 52 * always returns the correct integer provided it is 53 * representable. 54 * 55 * Constants : 56 * The hexadecimal values are the intended ones for the following 57 * constants. The decimal values may be used, provided that the 58 * compiler will convert from decimal to binary accurately enough 59 * to produce the hexadecimal values shown. 60 */ 61 62 #include <math.h> 63 64 #ifdef __STDC__ 65 static const double 66 #else 67 static double 68 #endif 69 bp[] = {1.0, 1.5,}, 70 dp_h[] = { 0.0, 5.84962487220764160156e-01,}, /* 0x3FE2B803, 0x40000000 */ 71 dp_l[] = { 0.0, 1.35003920212974897128e-08,}, /* 0x3E4CFDEB, 0x43CFD006 */ 72 zero = 0.0, 73 one = 1.0, 74 two = 2.0, 75 two53 = 9007199254740992.0, /* 0x43400000, 0x00000000 */ 76 huge = 1.0e300, 77 tiny = 1.0e-300, 78 /* poly coefs for (3/2)*(log(x)-2s-2/3*s**3 */ 79 L1 = 5.99999999999994648725e-01, /* 0x3FE33333, 0x33333303 */ 80 L2 = 4.28571428578550184252e-01, /* 0x3FDB6DB6, 0xDB6FABFF */ 81 L3 = 3.33333329818377432918e-01, /* 0x3FD55555, 0x518F264D */ 82 L4 = 2.72728123808534006489e-01, /* 0x3FD17460, 0xA91D4101 */ 83 L5 = 2.30660745775561754067e-01, /* 0x3FCD864A, 0x93C9DB65 */ 84 L6 = 2.06975017800338417784e-01, /* 0x3FCA7E28, 0x4A454EEF */ 85 P1 = 1.66666666666666019037e-01, /* 0x3FC55555, 0x5555553E */ 86 P2 = -2.77777777770155933842e-03, /* 0xBF66C16C, 0x16BEBD93 */ 87 P3 = 6.61375632143793436117e-05, /* 0x3F11566A, 0xAF25DE2C */ 88 P4 = -1.65339022054652515390e-06, /* 0xBEBBBD41, 0xC5D26BF1 */ 89 P5 = 4.13813679705723846039e-08, /* 0x3E663769, 0x72BEA4D0 */ 90 lg2 = 6.93147180559945286227e-01, /* 0x3FE62E42, 0xFEFA39EF */ 91 lg2_h = 6.93147182464599609375e-01, /* 0x3FE62E43, 0x00000000 */ 92 lg2_l = -1.90465429995776804525e-09, /* 0xBE205C61, 0x0CA86C39 */ 93 ovt = 8.0085662595372944372e-0017, /* -(1024-log2(ovfl+.5ulp)) */ 94 cp = 9.61796693925975554329e-01, /* 0x3FEEC709, 0xDC3A03FD =2/(3ln2) */ 95 cp_h = 9.61796700954437255859e-01, /* 0x3FEEC709, 0xE0000000 =(float)cp */ 96 cp_l = -7.02846165095275826516e-09, /* 0xBE3E2FE0, 0x145B01F5 =tail of cp_h*/ 97 ivln2 = 1.44269504088896338700e+00, /* 0x3FF71547, 0x652B82FE =1/ln2 */ 98 ivln2_h = 1.44269502162933349609e+00, /* 0x3FF71547, 0x60000000 =24b 1/ln2*/ 99 ivln2_l = 1.92596299112661746887e-08; /* 0x3E54AE0B, 0xF85DDF44 =1/ln2 tail*/ 100 101 #ifdef __STDC__ 102 double __ieee754_pow(double x, double y) 103 #else 104 double __ieee754_pow(x,y) 105 double x, y; 106 #endif 107 { 108 double z,ax,z_h,z_l,p_h,p_l; 109 double y1,t1,t2,r,s,t,u,v,w; 110 int i0,i1,i,j,k,yisint,n; 111 int hx,hy,ix,iy; 112 unsigned lx,ly; 113 114 i0 = ((*(int*)&one)>>29)^1; i1=1-i0; 115 hx = *(i0+(int*)&x); lx = *(i1+(int*)&x); 116 hy = *(i0+(int*)&y); ly = *(i1+(int*)&y); 117 ix = hx&0x7fffffff; iy = hy&0x7fffffff; 118 119 /* y==zero: x**0 = 1 */ 120 if((iy|ly)==0) return one; 121 122 /* +-NaN return x+y */ 123 if(ix > 0x7ff00000 || ((ix==0x7ff00000)&&(lx!=0)) || 124 iy > 0x7ff00000 || ((iy==0x7ff00000)&&(ly!=0))) 125 return x+y; 126 127 /* determine if y is an odd int when x < 0 128 * yisint = 0 ... y is not an integer 129 * yisint = 1 ... y is an odd int 130 * yisint = 2 ... y is an even int 131 */ 132 yisint = 0; 133 if(hx<0) { 134 if(iy>=0x43400000) yisint = 2; /* even integer y */ 135 else if(iy>=0x3ff00000) { 136 k = (iy>>20)-0x3ff; /* exponent */ 137 if(k>20) { 138 j = ly>>(52-k); 139 if((j<<(52-k))==ly) yisint = 2-(j&1); 140 } else if(ly==0) { 141 j = iy>>(20-k); 142 if((j<<(20-k))==iy) yisint = 2-(j&1); 143 } 144 } 145 } 146 147 /* special value of y */ 148 if(ly==0) { 149 if (iy==0x7ff00000) { /* y is +-inf */ 150 if(((ix-0x3ff00000)|lx)==0) 151 return y - y; /* inf**+-1 is NaN */ 152 else if (ix >= 0x3ff00000)/* (|x|>1)**+-inf = inf,0 */ 153 return (hy>=0)? y: zero; 154 else /* (|x|<1)**-,+inf = inf,0 */ 155 return (hy<0)?-y: zero; 156 } 157 if(iy==0x3ff00000) { /* y is +-1 */ 158 if(hy<0) return one/x; else return x; 159 } 160 if(hy==0x40000000) return x*x; /* y is 2 */ 161 if(hy==0x3fe00000) { /* y is 0.5 */ 162 if(hx>=0) /* x >= +0 */ 163 return sqrt(x); 164 } 165 } 166 167 ax = fabs(x); 168 /* special value of x */ 169 if(lx==0) { 170 if(ix==0x7ff00000||ix==0||ix==0x3ff00000){ 171 z = ax; /*x is +-0,+-inf,+-1*/ 172 if(hy<0) z = one/z; /* z = (1/|x|) */ 173 if(hx<0) { 174 if(((ix-0x3ff00000)|yisint)==0) { 175 z = (z-z)/(z-z); /* (-1)**non-int is NaN */ 176 } else if(yisint==1) 177 z = -z; /* (x<0)**odd = -(|x|**odd) */ 178 } 179 return z; 180 } 181 } 182 183 /* (x<0)**(non-int) is NaN */ 184 if((((hx>>31)+1)|yisint)==0) return (x-x)/(x-x); 185 186 /* |y| is huge */ 187 if(iy>0x41e00000) { /* if |y| > 2**31 */ 188 if(iy>0x43f00000){ /* if |y| > 2**64, must o/uflow */ 189 if(ix<=0x3fefffff) return (hy<0)? huge*huge:tiny*tiny; 190 if(ix>=0x3ff00000) return (hy>0)? huge*huge:tiny*tiny; 191 } 192 /* over/underflow if x is not close to one */ 193 if(ix<0x3fefffff) return (hy<0)? huge*huge:tiny*tiny; 194 if(ix>0x3ff00000) return (hy>0)? huge*huge:tiny*tiny; 195 /* now |1-x| is tiny <= 2**-20, suffice to compute 196 log(x) by x-x^2/2+x^3/3-x^4/4 */ 197 t = x-1; /* t has 20 trailing zeros */ 198 w = (t*t)*(0.5-t*(0.3333333333333333333333-t*0.25)); 199 u = ivln2_h*t; /* ivln2_h has 21 sig. bits */ 200 v = t*ivln2_l-w*ivln2; 201 t1 = u+v; 202 *(i1+(int*)&t1) = 0; 203 t2 = v-(t1-u); 204 } else { 205 double s2,s_h,s_l,t_h,t_l; 206 n = 0; 207 /* take care subnormal number */ 208 if(ix<0x00100000) 209 {ax *= two53; n -= 53; ix = *(i0+(int*)&ax); } 210 n += ((ix)>>20)-0x3ff; 211 j = ix&0x000fffff; 212 /* determine interval */ 213 ix = j|0x3ff00000; /* normalize ix */ 214 if(j<=0x3988E) k=0; /* |x|<sqrt(3/2) */ 215 else if(j<0xBB67A) k=1; /* |x|<sqrt(3) */ 216 else {k=0;n+=1;ix -= 0x00100000;} 217 *(i0+(int*)&ax) = ix; 218 219 /* compute s = s_h+s_l = (x-1)/(x+1) or (x-1.5)/(x+1.5) */ 220 u = ax-bp[k]; /* bp[0]=1.0, bp[1]=1.5 */ 221 v = one/(ax+bp[k]); 222 s = u*v; 223 s_h = s; 224 *(i1+(int*)&s_h) = 0; 225 /* t_h=ax+bp[k] High */ 226 t_h = zero; 227 *(i0+(int*)&t_h)=((ix>>1)|0x20000000)+0x00080000+(k<<18); 228 t_l = ax - (t_h-bp[k]); 229 s_l = v*((u-s_h*t_h)-s_h*t_l); 230 /* compute log(ax) */ 231 s2 = s*s; 232 r = s2*s2*(L1+s2*(L2+s2*(L3+s2*(L4+s2*(L5+s2*L6))))); 233 r += s_l*(s_h+s); 234 s2 = s_h*s_h; 235 t_h = 3.0+s2+r; 236 *(i1+(int*)&t_h) = 0; 237 t_l = r-((t_h-3.0)-s2); 238 /* u+v = s*(1+...) */ 239 u = s_h*t_h; 240 v = s_l*t_h+t_l*s; 241 /* 2/(3log2)*(s+...) */ 242 p_h = u+v; 243 *(i1+(int*)&p_h) = 0; 244 p_l = v-(p_h-u); 245 z_h = cp_h*p_h; /* cp_h+cp_l = 2/(3*log2) */ 246 z_l = cp_l*p_h+p_l*cp+dp_l[k]; 247 /* log2(ax) = (s+..)*2/(3*log2) = n + dp_h + z_h + z_l */ 248 t = (double)n; 249 t1 = (((z_h+z_l)+dp_h[k])+t); 250 *(i1+(int*)&t1) = 0; 251 t2 = z_l-(((t1-t)-dp_h[k])-z_h); 252 } 253 254 s = one; /* s (sign of result -ve**odd) = -1 else = 1 */ 255 if((((hx>>31)+1)|(yisint-1))==0) s = -one;/* (-ve)**(odd int) */ 256 257 /* split up y into y1+y2 and compute (y1+y2)*(t1+t2) */ 258 y1 = y; 259 *(i1+(int*)&y1) = 0; 260 p_l = (y-y1)*t1+y*t2; 261 p_h = y1*t1; 262 z = p_l+p_h; 263 j = *(i0+(int*)&z); 264 i = *(i1+(int*)&z); 265 if (j>=0x40900000) { /* z >= 1024 */ 266 if(((j-0x40900000)|i)!=0) /* if z > 1024 */ 267 return s*huge*huge; /* overflow */ 268 else { 269 if(p_l+ovt>z-p_h) return s*huge*huge; /* overflow */ 270 } 271 } else if((j&0x7fffffff)>=0x4090cc00 ) { /* z <= -1075 */ 272 if(((j-0xc090cc00)|i)!=0) /* z < -1075 */ 273 return s*tiny*tiny; /* underflow */ 274 else { 275 if(p_l<=z-p_h) return s*tiny*tiny; /* underflow */ 276 } 277 } 278 /* 279 * compute 2**(p_h+p_l) 280 */ 281 i = j&0x7fffffff; 282 k = (i>>20)-0x3ff; 283 n = 0; 284 if(i>0x3fe00000) { /* if |z| > 0.5, set n = [z+0.5] */ 285 n = j+(0x00100000>>(k+1)); 286 k = ((n&0x7fffffff)>>20)-0x3ff; /* new k for n */ 287 t = zero; 288 *(i0+(int*)&t) = (n&~(0x000fffff>>k)); 289 n = ((n&0x000fffff)|0x00100000)>>(20-k); 290 if(j<0) n = -n; 291 p_h -= t; 292 } 293 t = p_l+p_h; 294 *(i1+(int*)&t) = 0; 295 u = t*lg2_h; 296 v = (p_l-(t-p_h))*lg2+t*lg2_l; 297 z = u+v; 298 w = v-(z-u); 299 t = z*z; 300 t1 = z - t*(P1+t*(P2+t*(P3+t*(P4+t*P5)))); 301 r = (z*t1)/(t1-two)-(w+z*w); 302 z = one-(r-z); 303 j = *(i0+(int*)&z); 304 j += (n<<20); 305 if((j>>20)<=0) z = scalbn(z,n); /* subnormal output */ 306 else *(i0+(int*)&z) += (n<<20); 307 return s*z; 308 } 309