s_cbrt.c revision 153306
1210284Sjmallett/* @(#)s_cbrt.c 5.1 93/09/24 */
2232812Sjmallett/*
3215990Sjmallett * ====================================================
4210284Sjmallett * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
5210284Sjmallett *
6215990Sjmallett * Developed at SunPro, a Sun Microsystems, Inc. business.
7215990Sjmallett * Permission to use, copy, modify, and distribute this
8215990Sjmallett * software is freely granted, provided that this notice
9210284Sjmallett * is preserved.
10215990Sjmallett * ====================================================
11215990Sjmallett */
12210284Sjmallett
13215990Sjmallett#ifndef lint
14215990Sjmallettstatic char rcsid[] = "$FreeBSD: head/lib/msun/src/s_cbrt.c 153306 2005-12-11 19:51:30Z bde $";
15215990Sjmallett#endif
16215990Sjmallett
17215990Sjmallett#include "math.h"
18232812Sjmallett#include "math_private.h"
19215990Sjmallett
20215990Sjmallett/* cbrt(x)
21215990Sjmallett * Return cube root of x
22215990Sjmallett */
23215990Sjmallettstatic const u_int32_t
24215990Sjmallett	B1 = 715094163, /* B1 = (1023-1023/3-0.03306235651)*2**20 */
25215990Sjmallett	B2 = 696219795; /* B2 = (1023-1023/3-54/3-0.03306235651)*2**20 */
26215990Sjmallett
27215990Sjmallettstatic const double
28215990SjmallettC =  5.42857142857142815906e-01, /* 19/35     = 0x3FE15F15, 0xF15F15F1 */
29232812SjmallettD = -7.05306122448979611050e-01, /* -864/1225 = 0xBFE691DE, 0x2532C834 */
30215990SjmallettE =  1.41428571428571436819e+00, /* 99/70     = 0x3FF6A0EA, 0x0EA0EA0F */
31215990SjmallettF =  1.60714285714285720630e+00, /* 45/28     = 0x3FF9B6DB, 0x6DB6DB6E */
32215990SjmallettG =  3.57142857142857150787e-01; /* 5/14      = 0x3FD6DB6D, 0xB6DB6DB7 */
33215990Sjmallett
34215990Sjmallettdouble
35215990Sjmallettcbrt(double x)
36215990Sjmallett{
37215990Sjmallett	int32_t	hx;
38210284Sjmallett	double r,s,t=0.0,w;
39210284Sjmallett	u_int32_t sign;
40215990Sjmallett	u_int32_t high,low;
41210284Sjmallett
42210284Sjmallett	GET_HIGH_WORD(hx,x);
43210284Sjmallett	sign=hx&0x80000000; 		/* sign= sign(x) */
44210284Sjmallett	hx  ^=sign;
45210284Sjmallett	if(hx>=0x7ff00000) return(x+x); /* cbrt(NaN,INF) is itself */
46210284Sjmallett	GET_LOW_WORD(low,x);
47210284Sjmallett	if((hx|low)==0)
48210284Sjmallett	    return(x);		/* cbrt(0) is itself */
49232812Sjmallett
50210284Sjmallett	SET_HIGH_WORD(x,hx);	/* x <- |x| */
51210284Sjmallett    /*
52210284Sjmallett     * Rough cbrt to 5 bits:
53210284Sjmallett     *    cbrt(2**e*(1+m) ~= 2**(e/3)*(1+(e%3+m)/3)
54215990Sjmallett     * where e is integral and >= 0, m is real and in [0, 1), and "/" and
55215990Sjmallett     * "%" are integer division and modulus with rounding towards minus
56215990Sjmallett     * infinity.  The RHS is always >= the LHS and has a maximum relative
57232812Sjmallett     * error of about 1 in 16.  Adding a bias of -0.03306235651 to the
58215990Sjmallett     * (e%3+m)/3 term reduces the error to about 1 in 32. With the IEEE
59215990Sjmallett     * floating point representation, for finite positive normal values,
60215990Sjmallett     * ordinary integer divison of the value in bits magically gives
61215990Sjmallett     * almost exactly the RHS of the above provided we first subtract the
62215990Sjmallett     * exponent bias (1023 for doubles) and later add it back.  We do the
63232812Sjmallett     * subtraction virtually to keep e >= 0 so that ordinary integer
64232812Sjmallett     * division rounds towards minus infinity; this is also efficient.
65215990Sjmallett     */
66215990Sjmallett	if(hx<0x00100000) 		/* subnormal number */
67215990Sjmallett	  {SET_HIGH_WORD(t,0x43500000);	/* set t= 2**54 */
68232812Sjmallett	   t*=x; GET_HIGH_WORD(high,t); SET_HIGH_WORD(t,high/3+B2);
69232812Sjmallett	  }
70215990Sjmallett	else
71215990Sjmallett	  SET_HIGH_WORD(t,hx/3+B1);
72215990Sjmallett
73215990Sjmallett    /* new cbrt to 23 bits; may be implemented in single precision */
74215990Sjmallett	r=t*t/x;
75215990Sjmallett	s=C+r*t;
76215990Sjmallett	t*=G+F/(s+E+D/s);
77215990Sjmallett
78215990Sjmallett    /* chop t to 20 bits and make it larger than cbrt(x) */
79215990Sjmallett	GET_HIGH_WORD(high,t);
80215990Sjmallett	INSERT_WORDS(t,high+0x00000001,0);
81215990Sjmallett
82215990Sjmallett    /* one step Newton iteration to 53 bits with error less than 0.667 ulps */
83215990Sjmallett	s=t*t;		/* t*t is exact */
84215990Sjmallett	r=x/s;
85215990Sjmallett	w=t+t;
86215990Sjmallett	r=(r-t)/(w+r);	/* r-t is exact */
87215990Sjmallett	t=t+t*r;
88215990Sjmallett
89215990Sjmallett    /* restore the sign bit */
90215990Sjmallett	GET_HIGH_WORD(high,t);
91215990Sjmallett	SET_HIGH_WORD(t,high|sign);
92215990Sjmallett	return(t);
93215990Sjmallett}
94215990Sjmallett