xref: /freebsd/lib/msun/ld128/e_rem_pio2l.h (revision a466cc55373fc3cf86837f09da729535b57e69a1)
1 /* From: @(#)e_rem_pio2.c 1.4 95/01/18 */
2 /*
3  * ====================================================
4  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
5  * Copyright (c) 2008 Steven G. Kargl, David Schultz, Bruce D. Evans.
6  *
7  * Developed at SunSoft, a Sun Microsystems, Inc. business.
8  * Permission to use, copy, modify, and distribute this
9  * software is freely granted, provided that this notice
10  * is preserved.
11  * ====================================================
12  *
13  * Optimized by Bruce D. Evans.
14  */
15 
16 #include <sys/cdefs.h>
17 __FBSDID("$FreeBSD$");
18 
19 /* ld128 version of __ieee754_rem_pio2l(x,y)
20  *
21  * return the remainder of x rem pi/2 in y[0]+y[1]
22  * use __kernel_rem_pio2()
23  */
24 
25 #include <float.h>
26 
27 #include "math.h"
28 #include "math_private.h"
29 #include "fpmath.h"
30 
31 #define	BIAS	(LDBL_MAX_EXP - 1)
32 
33 /*
34  * XXX need to verify that nonzero integer multiples of pi/2 within the
35  * range get no closer to a long double than 2**-140, or that
36  * ilogb(x) + ilogb(min_delta) < 45 - -140.
37  */
38 /*
39  * invpio2:  113 bits of 2/pi
40  * pio2_1:   first  68 bits of pi/2
41  * pio2_1t:  pi/2 - pio2_1
42  * pio2_2:   second 68 bits of pi/2
43  * pio2_2t:  pi/2 - (pio2_1+pio2_2)
44  * pio2_3:   third  68 bits of pi/2
45  * pio2_3t:  pi/2 - (pio2_1+pio2_2+pio2_3)
46  */
47 
48 static const double
49 zero =  0.00000000000000000000e+00, /* 0x00000000, 0x00000000 */
50 two24 =  1.67772160000000000000e+07; /* 0x41700000, 0x00000000 */
51 
52 static const long double
53 invpio2 =  6.3661977236758134307553505349005747e-01L,	/*  0x145f306dc9c882a53f84eafa3ea6a.0p-113 */
54 pio2_1  =  1.5707963267948966192292994253909555e+00L,	/*  0x1921fb54442d18469800000000000.0p-112 */
55 pio2_1t =  2.0222662487959507323996846200947577e-21L,	/*  0x13198a2e03707344a4093822299f3.0p-181 */
56 pio2_2  =  2.0222662487959507323994779168837751e-21L,	/*  0x13198a2e03707344a400000000000.0p-181 */
57 pio2_2t =  2.0670321098263988236496903051604844e-43L,	/*  0x127044533e63a0105df531d89cd91.0p-254 */
58 pio2_3  =  2.0670321098263988236499468110329591e-43L,	/*  0x127044533e63a0105e00000000000.0p-254 */
59 pio2_3t = -2.5650587247459238361625433492959285e-65L;	/* -0x159c4ec64ddaeb5f78671cbfb2210.0p-327 */
60 
61 static inline __always_inline int
62 __ieee754_rem_pio2l(long double x, long double *y)
63 {
64 	union IEEEl2bits u,u1;
65 	long double z,w,t,r,fn;
66 	double tx[5],ty[3];
67 	int64_t n;
68 	int e0,ex,i,j,nx;
69 	int16_t expsign;
70 
71 	u.e = x;
72 	expsign = u.xbits.expsign;
73 	ex = expsign & 0x7fff;
74 	if (ex < BIAS + 45 || ex == BIAS + 45 &&
75 	    u.bits.manh < 0x921fb54442d1LL) {
76 	    /* |x| ~< 2^45*(pi/2), medium size */
77 	    /* TODO: use only double precision for fn, as in expl(). */
78 	    fn = rnintl(x * invpio2);
79 	    n  = i64rint(fn);
80 	    r  = x-fn*pio2_1;
81 	    w  = fn*pio2_1t;	/* 1st round good to 180 bit */
82 	    {
83 		union IEEEl2bits u2;
84 	        int ex1;
85 	        j  = ex;
86 	        y[0] = r-w;
87 		u2.e = y[0];
88 		ex1 = u2.xbits.expsign & 0x7fff;
89 	        i = j-ex1;
90 	        if(i>51) {  /* 2nd iteration needed, good to 248 */
91 		    t  = r;
92 		    w  = fn*pio2_2;
93 		    r  = t-w;
94 		    w  = fn*pio2_2t-((t-r)-w);
95 		    y[0] = r-w;
96 		    u2.e = y[0];
97 		    ex1 = u2.xbits.expsign & 0x7fff;
98 		    i = j-ex1;
99 		    if(i>119) {	/* 3rd iteration need, 316 bits acc */
100 		    	t  = r;	/* will cover all possible cases */
101 		    	w  = fn*pio2_3;
102 		    	r  = t-w;
103 		    	w  = fn*pio2_3t-((t-r)-w);
104 		    	y[0] = r-w;
105 		    }
106 		}
107 	    }
108 	    y[1] = (r-y[0])-w;
109 	    return n;
110 	}
111     /*
112      * all other (large) arguments
113      */
114 	if(ex==0x7fff) {		/* x is inf or NaN */
115 	    y[0]=y[1]=x-x; return 0;
116 	}
117     /* set z = scalbn(|x|,ilogb(x)-23) */
118 	u1.e = x;
119 	e0 = ex - BIAS - 23;		/* e0 = ilogb(|x|)-23; */
120 	u1.xbits.expsign = ex - e0;
121 	z = u1.e;
122 	for(i=0;i<4;i++) {
123 		tx[i] = (double)((int32_t)(z));
124 		z     = (z-tx[i])*two24;
125 	}
126 	tx[4] = z;
127 	nx = 5;
128 	while(tx[nx-1]==zero) nx--;	/* skip zero term */
129 	n  =  __kernel_rem_pio2(tx,ty,e0,nx,3);
130 	t = (long double)ty[2] + ty[1];
131 	r = t + ty[0];
132 	w = ty[0] - (r - t);
133 	if(expsign<0) {y[0] = -r; y[1] = -w; return -n;}
134 	y[0] = r; y[1] = w; return n;
135 }
136