xref: /freebsd/lib/msun/ld128/e_rem_pio2l.h (revision 06c3fb2749bda94cb5201f81ffdb8fa6c3161b2e)
1 /*
2  * ====================================================
3  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
4  * Copyright (c) 2008 Steven G. Kargl, David Schultz, Bruce D. Evans.
5  *
6  * Developed at SunSoft, 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  * Optimized by Bruce D. Evans.
13  */
14 
15 #include <sys/cdefs.h>
16 /* ld128 version of __ieee754_rem_pio2l(x,y)
17  *
18  * return the remainder of x rem pi/2 in y[0]+y[1]
19  * use __kernel_rem_pio2()
20  */
21 
22 #include <float.h>
23 
24 #include "math.h"
25 #include "math_private.h"
26 #include "fpmath.h"
27 
28 #define	BIAS	(LDBL_MAX_EXP - 1)
29 
30 /*
31  * XXX need to verify that nonzero integer multiples of pi/2 within the
32  * range get no closer to a long double than 2**-140, or that
33  * ilogb(x) + ilogb(min_delta) < 45 - -140.
34  */
35 /*
36  * invpio2:  113 bits of 2/pi
37  * pio2_1:   first  68 bits of pi/2
38  * pio2_1t:  pi/2 - pio2_1
39  * pio2_2:   second 68 bits of pi/2
40  * pio2_2t:  pi/2 - (pio2_1+pio2_2)
41  * pio2_3:   third  68 bits of pi/2
42  * pio2_3t:  pi/2 - (pio2_1+pio2_2+pio2_3)
43  */
44 
45 static const double
46 zero =  0.00000000000000000000e+00, /* 0x00000000, 0x00000000 */
47 two24 =  1.67772160000000000000e+07; /* 0x41700000, 0x00000000 */
48 
49 static const long double
50 invpio2 =  6.3661977236758134307553505349005747e-01L,	/*  0x145f306dc9c882a53f84eafa3ea6a.0p-113 */
51 pio2_1  =  1.5707963267948966192292994253909555e+00L,	/*  0x1921fb54442d18469800000000000.0p-112 */
52 pio2_1t =  2.0222662487959507323996846200947577e-21L,	/*  0x13198a2e03707344a4093822299f3.0p-181 */
53 pio2_2  =  2.0222662487959507323994779168837751e-21L,	/*  0x13198a2e03707344a400000000000.0p-181 */
54 pio2_2t =  2.0670321098263988236496903051604844e-43L,	/*  0x127044533e63a0105df531d89cd91.0p-254 */
55 pio2_3  =  2.0670321098263988236499468110329591e-43L,	/*  0x127044533e63a0105e00000000000.0p-254 */
56 pio2_3t = -2.5650587247459238361625433492959285e-65L;	/* -0x159c4ec64ddaeb5f78671cbfb2210.0p-327 */
57 
58 static inline __always_inline int
59 __ieee754_rem_pio2l(long double x, long double *y)
60 {
61 	union IEEEl2bits u,u1;
62 	long double z,w,t,r,fn;
63 	double tx[5],ty[3];
64 	int64_t n;
65 	int e0,ex,i,j,nx;
66 	int16_t expsign;
67 
68 	u.e = x;
69 	expsign = u.xbits.expsign;
70 	ex = expsign & 0x7fff;
71 	if (ex < BIAS + 45 || ex == BIAS + 45 &&
72 	    u.bits.manh < 0x921fb54442d1LL) {
73 	    /* |x| ~< 2^45*(pi/2), medium size */
74 	    /* TODO: use only double precision for fn, as in expl(). */
75 	    fn = rnintl(x * invpio2);
76 	    n  = i64rint(fn);
77 	    r  = x-fn*pio2_1;
78 	    w  = fn*pio2_1t;	/* 1st round good to 180 bit */
79 	    {
80 		union IEEEl2bits u2;
81 	        int ex1;
82 	        j  = ex;
83 	        y[0] = r-w;
84 		u2.e = y[0];
85 		ex1 = u2.xbits.expsign & 0x7fff;
86 	        i = j-ex1;
87 	        if(i>51) {  /* 2nd iteration needed, good to 248 */
88 		    t  = r;
89 		    w  = fn*pio2_2;
90 		    r  = t-w;
91 		    w  = fn*pio2_2t-((t-r)-w);
92 		    y[0] = r-w;
93 		    u2.e = y[0];
94 		    ex1 = u2.xbits.expsign & 0x7fff;
95 		    i = j-ex1;
96 		    if(i>119) {	/* 3rd iteration need, 316 bits acc */
97 		    	t  = r;	/* will cover all possible cases */
98 		    	w  = fn*pio2_3;
99 		    	r  = t-w;
100 		    	w  = fn*pio2_3t-((t-r)-w);
101 		    	y[0] = r-w;
102 		    }
103 		}
104 	    }
105 	    y[1] = (r-y[0])-w;
106 	    return n;
107 	}
108     /*
109      * all other (large) arguments
110      */
111 	if(ex==0x7fff) {		/* x is inf or NaN */
112 	    y[0]=y[1]=x-x; return 0;
113 	}
114     /* set z = scalbn(|x|,ilogb(x)-23) */
115 	u1.e = x;
116 	e0 = ex - BIAS - 23;		/* e0 = ilogb(|x|)-23; */
117 	u1.xbits.expsign = ex - e0;
118 	z = u1.e;
119 	for(i=0;i<4;i++) {
120 		tx[i] = (double)((int32_t)(z));
121 		z     = (z-tx[i])*two24;
122 	}
123 	tx[4] = z;
124 	nx = 5;
125 	while(tx[nx-1]==zero) nx--;	/* skip zero term */
126 	n  =  __kernel_rem_pio2(tx,ty,e0,nx,3);
127 	t = (long double)ty[2] + ty[1];
128 	r = t + ty[0];
129 	w = ty[0] - (r - t);
130 	if(expsign<0) {y[0] = -r; y[1] = -w; return -n;}
131 	y[0] = r; y[1] = w; return n;
132 }
133