xref: /freebsd/lib/msun/src/e_pow.c (revision 9336e0699bda8a301cd2bfa37106b6ec5e32012e)
1 /* @(#)e_pow.c 1.5 04/04/22 SMI */
2 /*
3  * ====================================================
4  * Copyright (C) 2004 by Sun Microsystems, Inc. All rights reserved.
5  *
6  * Permission to use, copy, modify, and distribute this
7  * software is freely granted, provided that this notice
8  * is preserved.
9  * ====================================================
10  */
11 
12 #ifndef lint
13 static char rcsid[] = "$FreeBSD$";
14 #endif
15 
16 /* __ieee754_pow(x,y) return x**y
17  *
18  *		      n
19  * Method:  Let x =  2   * (1+f)
20  *	1. Compute and return log2(x) in two pieces:
21  *		log2(x) = w1 + w2,
22  *	   where w1 has 53-24 = 29 bit trailing zeros.
23  *	2. Perform y*log2(x) = n+y' by simulating muti-precision
24  *	   arithmetic, where |y'|<=0.5.
25  *	3. Return x**y = 2**n*exp(y'*log2)
26  *
27  * Special cases:
28  *	1.  (anything) ** 0  is 1
29  *	2.  (anything) ** 1  is itself
30  *	3.  (anything) ** NAN is NAN
31  *	4.  NAN ** (anything except 0) is NAN
32  *	5.  +-(|x| > 1) **  +INF is +INF
33  *	6.  +-(|x| > 1) **  -INF is +0
34  *	7.  +-(|x| < 1) **  +INF is +0
35  *	8.  +-(|x| < 1) **  -INF is +INF
36  *	9.  +-1         ** +-INF is NAN
37  *	10. +0 ** (+anything except 0, NAN)               is +0
38  *	11. -0 ** (+anything except 0, NAN, odd integer)  is +0
39  *	12. +0 ** (-anything except 0, NAN)               is +INF
40  *	13. -0 ** (-anything except 0, NAN, odd integer)  is +INF
41  *	14. -0 ** (odd integer) = -( +0 ** (odd integer) )
42  *	15. +INF ** (+anything except 0,NAN) is +INF
43  *	16. +INF ** (-anything except 0,NAN) is +0
44  *	17. -INF ** (anything)  = -0 ** (-anything)
45  *	18. (-anything) ** (integer) is (-1)**(integer)*(+anything**integer)
46  *	19. (-anything except 0 and inf) ** (non-integer) is NAN
47  *
48  * Accuracy:
49  *	pow(x,y) returns x**y nearly rounded. In particular
50  *			pow(integer,integer)
51  *	always returns the correct integer provided it is
52  *	representable.
53  *
54  * Constants :
55  * The hexadecimal values are the intended ones for the following
56  * constants. The decimal values may be used, provided that the
57  * compiler will convert from decimal to binary accurately enough
58  * to produce the hexadecimal values shown.
59  */
60 
61 #include "math.h"
62 #include "math_private.h"
63 
64 static const double
65 bp[] = {1.0, 1.5,},
66 dp_h[] = { 0.0, 5.84962487220764160156e-01,}, /* 0x3FE2B803, 0x40000000 */
67 dp_l[] = { 0.0, 1.35003920212974897128e-08,}, /* 0x3E4CFDEB, 0x43CFD006 */
68 zero    =  0.0,
69 one	=  1.0,
70 two	=  2.0,
71 two53	=  9007199254740992.0,	/* 0x43400000, 0x00000000 */
72 huge	=  1.0e300,
73 tiny    =  1.0e-300,
74 	/* poly coefs for (3/2)*(log(x)-2s-2/3*s**3 */
75 L1  =  5.99999999999994648725e-01, /* 0x3FE33333, 0x33333303 */
76 L2  =  4.28571428578550184252e-01, /* 0x3FDB6DB6, 0xDB6FABFF */
77 L3  =  3.33333329818377432918e-01, /* 0x3FD55555, 0x518F264D */
78 L4  =  2.72728123808534006489e-01, /* 0x3FD17460, 0xA91D4101 */
79 L5  =  2.30660745775561754067e-01, /* 0x3FCD864A, 0x93C9DB65 */
80 L6  =  2.06975017800338417784e-01, /* 0x3FCA7E28, 0x4A454EEF */
81 P1   =  1.66666666666666019037e-01, /* 0x3FC55555, 0x5555553E */
82 P2   = -2.77777777770155933842e-03, /* 0xBF66C16C, 0x16BEBD93 */
83 P3   =  6.61375632143793436117e-05, /* 0x3F11566A, 0xAF25DE2C */
84 P4   = -1.65339022054652515390e-06, /* 0xBEBBBD41, 0xC5D26BF1 */
85 P5   =  4.13813679705723846039e-08, /* 0x3E663769, 0x72BEA4D0 */
86 lg2  =  6.93147180559945286227e-01, /* 0x3FE62E42, 0xFEFA39EF */
87 lg2_h  =  6.93147182464599609375e-01, /* 0x3FE62E43, 0x00000000 */
88 lg2_l  = -1.90465429995776804525e-09, /* 0xBE205C61, 0x0CA86C39 */
89 ovt =  8.0085662595372944372e-0017, /* -(1024-log2(ovfl+.5ulp)) */
90 cp    =  9.61796693925975554329e-01, /* 0x3FEEC709, 0xDC3A03FD =2/(3ln2) */
91 cp_h  =  9.61796700954437255859e-01, /* 0x3FEEC709, 0xE0000000 =(float)cp */
92 cp_l  = -7.02846165095275826516e-09, /* 0xBE3E2FE0, 0x145B01F5 =tail of cp_h*/
93 ivln2    =  1.44269504088896338700e+00, /* 0x3FF71547, 0x652B82FE =1/ln2 */
94 ivln2_h  =  1.44269502162933349609e+00, /* 0x3FF71547, 0x60000000 =24b 1/ln2*/
95 ivln2_l  =  1.92596299112661746887e-08; /* 0x3E54AE0B, 0xF85DDF44 =1/ln2 tail*/
96 
97 double
98 __ieee754_pow(double x, double y)
99 {
100 	double z,ax,z_h,z_l,p_h,p_l;
101 	double y1,t1,t2,r,s,t,u,v,w;
102 	int32_t i,j,k,yisint,n;
103 	int32_t hx,hy,ix,iy;
104 	u_int32_t lx,ly;
105 
106 	EXTRACT_WORDS(hx,lx,x);
107 	EXTRACT_WORDS(hy,ly,y);
108 	ix = hx&0x7fffffff;  iy = hy&0x7fffffff;
109 
110     /* y==zero: x**0 = 1 */
111 	if((iy|ly)==0) return one;
112 
113     /* +-NaN return x+y */
114 	if(ix > 0x7ff00000 || ((ix==0x7ff00000)&&(lx!=0)) ||
115 	   iy > 0x7ff00000 || ((iy==0x7ff00000)&&(ly!=0)))
116 		return x+y;
117 
118     /* determine if y is an odd int when x < 0
119      * yisint = 0	... y is not an integer
120      * yisint = 1	... y is an odd int
121      * yisint = 2	... y is an even int
122      */
123 	yisint  = 0;
124 	if(hx<0) {
125 	    if(iy>=0x43400000) yisint = 2; /* even integer y */
126 	    else if(iy>=0x3ff00000) {
127 		k = (iy>>20)-0x3ff;	   /* exponent */
128 		if(k>20) {
129 		    j = ly>>(52-k);
130 		    if((j<<(52-k))==ly) yisint = 2-(j&1);
131 		} else if(ly==0) {
132 		    j = iy>>(20-k);
133 		    if((j<<(20-k))==iy) yisint = 2-(j&1);
134 		}
135 	    }
136 	}
137 
138     /* special value of y */
139 	if(ly==0) {
140 	    if (iy==0x7ff00000) {	/* y is +-inf */
141 	        if(((ix-0x3ff00000)|lx)==0)
142 		    return  y - y;	/* inf**+-1 is NaN */
143 	        else if (ix >= 0x3ff00000)/* (|x|>1)**+-inf = inf,0 */
144 		    return (hy>=0)? y: zero;
145 	        else			/* (|x|<1)**-,+inf = inf,0 */
146 		    return (hy<0)?-y: zero;
147 	    }
148 	    if(iy==0x3ff00000) {	/* y is  +-1 */
149 		if(hy<0) return one/x; else return x;
150 	    }
151 	    if(hy==0x40000000) return x*x; /* y is  2 */
152 	    if(hy==0x3fe00000) {	/* y is  0.5 */
153 		if(hx>=0)	/* x >= +0 */
154 		return sqrt(x);
155 	    }
156 	}
157 
158 	ax   = fabs(x);
159     /* special value of x */
160 	if(lx==0) {
161 	    if(ix==0x7ff00000||ix==0||ix==0x3ff00000){
162 		z = ax;			/*x is +-0,+-inf,+-1*/
163 		if(hy<0) z = one/z;	/* z = (1/|x|) */
164 		if(hx<0) {
165 		    if(((ix-0x3ff00000)|yisint)==0) {
166 			z = (z-z)/(z-z); /* (-1)**non-int is NaN */
167 		    } else if(yisint==1)
168 			z = -z;		/* (x<0)**odd = -(|x|**odd) */
169 		}
170 		return z;
171 	    }
172 	}
173 
174     /* CYGNUS LOCAL + fdlibm-5.3 fix: This used to be
175 	n = (hx>>31)+1;
176        but ANSI C says a right shift of a signed negative quantity is
177        implementation defined.  */
178 	n = ((u_int32_t)hx>>31)-1;
179 
180     /* (x<0)**(non-int) is NaN */
181 	if((n|yisint)==0) return (x-x)/(x-x);
182 
183 	s = one; /* s (sign of result -ve**odd) = -1 else = 1 */
184 	if((n|(yisint-1))==0) s = -one;/* (-ve)**(odd int) */
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)? s*huge*huge:s*tiny*tiny;
194 	    if(ix>0x3ff00000) return (hy>0)? s*huge*huge:s*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 = ax-one;		/* 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 	    SET_LOW_WORD(t1,0);
203 	    t2 = v-(t1-u);
204 	} else {
205 	    double ss,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; GET_HIGH_WORD(ix,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 	    SET_HIGH_WORD(ax,ix);
218 
219 	/* compute ss = 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 	    ss = u*v;
223 	    s_h = ss;
224 	    SET_LOW_WORD(s_h,0);
225 	/* t_h=ax+bp[k] High */
226 	    t_h = zero;
227 	    SET_HIGH_WORD(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 = ss*ss;
232 	    r = s2*s2*(L1+s2*(L2+s2*(L3+s2*(L4+s2*(L5+s2*L6)))));
233 	    r += s_l*(s_h+ss);
234 	    s2  = s_h*s_h;
235 	    t_h = 3.0+s2+r;
236 	    SET_LOW_WORD(t_h,0);
237 	    t_l = r-((t_h-3.0)-s2);
238 	/* u+v = ss*(1+...) */
239 	    u = s_h*t_h;
240 	    v = s_l*t_h+t_l*ss;
241 	/* 2/(3log2)*(ss+...) */
242 	    p_h = u+v;
243 	    SET_LOW_WORD(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) = (ss+..)*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 	    SET_LOW_WORD(t1,0);
251 	    t2 = z_l-(((t1-t)-dp_h[k])-z_h);
252 	}
253 
254     /* split up y into y1+y2 and compute (y1+y2)*(t1+t2) */
255 	y1  = y;
256 	SET_LOW_WORD(y1,0);
257 	p_l = (y-y1)*t1+y*t2;
258 	p_h = y1*t1;
259 	z = p_l+p_h;
260 	EXTRACT_WORDS(j,i,z);
261 	if (j>=0x40900000) {				/* z >= 1024 */
262 	    if(((j-0x40900000)|i)!=0)			/* if z > 1024 */
263 		return s*huge*huge;			/* overflow */
264 	    else {
265 		if(p_l+ovt>z-p_h) return s*huge*huge;	/* overflow */
266 	    }
267 	} else if((j&0x7fffffff)>=0x4090cc00 ) {	/* z <= -1075 */
268 	    if(((j-0xc090cc00)|i)!=0) 		/* z < -1075 */
269 		return s*tiny*tiny;		/* underflow */
270 	    else {
271 		if(p_l<=z-p_h) return s*tiny*tiny;	/* underflow */
272 	    }
273 	}
274     /*
275      * compute 2**(p_h+p_l)
276      */
277 	i = j&0x7fffffff;
278 	k = (i>>20)-0x3ff;
279 	n = 0;
280 	if(i>0x3fe00000) {		/* if |z| > 0.5, set n = [z+0.5] */
281 	    n = j+(0x00100000>>(k+1));
282 	    k = ((n&0x7fffffff)>>20)-0x3ff;	/* new k for n */
283 	    t = zero;
284 	    SET_HIGH_WORD(t,n&~(0x000fffff>>k));
285 	    n = ((n&0x000fffff)|0x00100000)>>(20-k);
286 	    if(j<0) n = -n;
287 	    p_h -= t;
288 	}
289 	t = p_l+p_h;
290 	SET_LOW_WORD(t,0);
291 	u = t*lg2_h;
292 	v = (p_l-(t-p_h))*lg2+t*lg2_l;
293 	z = u+v;
294 	w = v-(z-u);
295 	t  = z*z;
296 	t1  = z - t*(P1+t*(P2+t*(P3+t*(P4+t*P5))));
297 	r  = (z*t1)/(t1-two)-(w+z*w);
298 	z  = one-(r-z);
299 	GET_HIGH_WORD(j,z);
300 	j += (n<<20);
301 	if((j>>20)<=0) z = scalbn(z,n);	/* subnormal output */
302 	else SET_HIGH_WORD(z,j);
303 	return s*z;
304 }
305