1*f504f610SAugustin Cavalier /* origin: FreeBSD /usr/src/lib/msun/src/e_rem_pio2.c */
2*f504f610SAugustin Cavalier /*
3*f504f610SAugustin Cavalier * ====================================================
4*f504f610SAugustin Cavalier * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
5*f504f610SAugustin Cavalier *
6*f504f610SAugustin Cavalier * Developed at SunSoft, a Sun Microsystems, Inc. business.
7*f504f610SAugustin Cavalier * Permission to use, copy, modify, and distribute this
8*f504f610SAugustin Cavalier * software is freely granted, provided that this notice
9*f504f610SAugustin Cavalier * is preserved.
10*f504f610SAugustin Cavalier * ====================================================
11*f504f610SAugustin Cavalier *
12*f504f610SAugustin Cavalier * Optimized by Bruce D. Evans.
13*f504f610SAugustin Cavalier */
14*f504f610SAugustin Cavalier /* __rem_pio2(x,y)
15*f504f610SAugustin Cavalier *
16*f504f610SAugustin Cavalier * return the remainder of x rem pi/2 in y[0]+y[1]
17*f504f610SAugustin Cavalier * use __rem_pio2_large() for large x
18*f504f610SAugustin Cavalier */
19*f504f610SAugustin Cavalier
20*f504f610SAugustin Cavalier #include "libm.h"
21*f504f610SAugustin Cavalier
22*f504f610SAugustin Cavalier #if FLT_EVAL_METHOD==0 || FLT_EVAL_METHOD==1
23*f504f610SAugustin Cavalier #define EPS DBL_EPSILON
24*f504f610SAugustin Cavalier #elif FLT_EVAL_METHOD==2
25*f504f610SAugustin Cavalier #define EPS LDBL_EPSILON
26*f504f610SAugustin Cavalier #endif
27*f504f610SAugustin Cavalier
28*f504f610SAugustin Cavalier /*
29*f504f610SAugustin Cavalier * invpio2: 53 bits of 2/pi
30*f504f610SAugustin Cavalier * pio2_1: first 33 bit of pi/2
31*f504f610SAugustin Cavalier * pio2_1t: pi/2 - pio2_1
32*f504f610SAugustin Cavalier * pio2_2: second 33 bit of pi/2
33*f504f610SAugustin Cavalier * pio2_2t: pi/2 - (pio2_1+pio2_2)
34*f504f610SAugustin Cavalier * pio2_3: third 33 bit of pi/2
35*f504f610SAugustin Cavalier * pio2_3t: pi/2 - (pio2_1+pio2_2+pio2_3)
36*f504f610SAugustin Cavalier */
37*f504f610SAugustin Cavalier static const double
38*f504f610SAugustin Cavalier toint = 1.5/EPS,
39*f504f610SAugustin Cavalier invpio2 = 6.36619772367581382433e-01, /* 0x3FE45F30, 0x6DC9C883 */
40*f504f610SAugustin Cavalier pio2_1 = 1.57079632673412561417e+00, /* 0x3FF921FB, 0x54400000 */
41*f504f610SAugustin Cavalier pio2_1t = 6.07710050650619224932e-11, /* 0x3DD0B461, 0x1A626331 */
42*f504f610SAugustin Cavalier pio2_2 = 6.07710050630396597660e-11, /* 0x3DD0B461, 0x1A600000 */
43*f504f610SAugustin Cavalier pio2_2t = 2.02226624879595063154e-21, /* 0x3BA3198A, 0x2E037073 */
44*f504f610SAugustin Cavalier pio2_3 = 2.02226624871116645580e-21, /* 0x3BA3198A, 0x2E000000 */
45*f504f610SAugustin Cavalier pio2_3t = 8.47842766036889956997e-32; /* 0x397B839A, 0x252049C1 */
46*f504f610SAugustin Cavalier
47*f504f610SAugustin Cavalier /* caller must handle the case when reduction is not needed: |x| ~<= pi/4 */
__rem_pio2(double x,double * y)48*f504f610SAugustin Cavalier int __rem_pio2(double x, double *y)
49*f504f610SAugustin Cavalier {
50*f504f610SAugustin Cavalier union {double f; uint64_t i;} u = {x};
51*f504f610SAugustin Cavalier double_t z,w,t,r,fn;
52*f504f610SAugustin Cavalier double tx[3],ty[2];
53*f504f610SAugustin Cavalier uint32_t ix;
54*f504f610SAugustin Cavalier int sign, n, ex, ey, i;
55*f504f610SAugustin Cavalier
56*f504f610SAugustin Cavalier sign = u.i>>63;
57*f504f610SAugustin Cavalier ix = u.i>>32 & 0x7fffffff;
58*f504f610SAugustin Cavalier if (ix <= 0x400f6a7a) { /* |x| ~<= 5pi/4 */
59*f504f610SAugustin Cavalier if ((ix & 0xfffff) == 0x921fb) /* |x| ~= pi/2 or 2pi/2 */
60*f504f610SAugustin Cavalier goto medium; /* cancellation -- use medium case */
61*f504f610SAugustin Cavalier if (ix <= 0x4002d97c) { /* |x| ~<= 3pi/4 */
62*f504f610SAugustin Cavalier if (!sign) {
63*f504f610SAugustin Cavalier z = x - pio2_1; /* one round good to 85 bits */
64*f504f610SAugustin Cavalier y[0] = z - pio2_1t;
65*f504f610SAugustin Cavalier y[1] = (z-y[0]) - pio2_1t;
66*f504f610SAugustin Cavalier return 1;
67*f504f610SAugustin Cavalier } else {
68*f504f610SAugustin Cavalier z = x + pio2_1;
69*f504f610SAugustin Cavalier y[0] = z + pio2_1t;
70*f504f610SAugustin Cavalier y[1] = (z-y[0]) + pio2_1t;
71*f504f610SAugustin Cavalier return -1;
72*f504f610SAugustin Cavalier }
73*f504f610SAugustin Cavalier } else {
74*f504f610SAugustin Cavalier if (!sign) {
75*f504f610SAugustin Cavalier z = x - 2*pio2_1;
76*f504f610SAugustin Cavalier y[0] = z - 2*pio2_1t;
77*f504f610SAugustin Cavalier y[1] = (z-y[0]) - 2*pio2_1t;
78*f504f610SAugustin Cavalier return 2;
79*f504f610SAugustin Cavalier } else {
80*f504f610SAugustin Cavalier z = x + 2*pio2_1;
81*f504f610SAugustin Cavalier y[0] = z + 2*pio2_1t;
82*f504f610SAugustin Cavalier y[1] = (z-y[0]) + 2*pio2_1t;
83*f504f610SAugustin Cavalier return -2;
84*f504f610SAugustin Cavalier }
85*f504f610SAugustin Cavalier }
86*f504f610SAugustin Cavalier }
87*f504f610SAugustin Cavalier if (ix <= 0x401c463b) { /* |x| ~<= 9pi/4 */
88*f504f610SAugustin Cavalier if (ix <= 0x4015fdbc) { /* |x| ~<= 7pi/4 */
89*f504f610SAugustin Cavalier if (ix == 0x4012d97c) /* |x| ~= 3pi/2 */
90*f504f610SAugustin Cavalier goto medium;
91*f504f610SAugustin Cavalier if (!sign) {
92*f504f610SAugustin Cavalier z = x - 3*pio2_1;
93*f504f610SAugustin Cavalier y[0] = z - 3*pio2_1t;
94*f504f610SAugustin Cavalier y[1] = (z-y[0]) - 3*pio2_1t;
95*f504f610SAugustin Cavalier return 3;
96*f504f610SAugustin Cavalier } else {
97*f504f610SAugustin Cavalier z = x + 3*pio2_1;
98*f504f610SAugustin Cavalier y[0] = z + 3*pio2_1t;
99*f504f610SAugustin Cavalier y[1] = (z-y[0]) + 3*pio2_1t;
100*f504f610SAugustin Cavalier return -3;
101*f504f610SAugustin Cavalier }
102*f504f610SAugustin Cavalier } else {
103*f504f610SAugustin Cavalier if (ix == 0x401921fb) /* |x| ~= 4pi/2 */
104*f504f610SAugustin Cavalier goto medium;
105*f504f610SAugustin Cavalier if (!sign) {
106*f504f610SAugustin Cavalier z = x - 4*pio2_1;
107*f504f610SAugustin Cavalier y[0] = z - 4*pio2_1t;
108*f504f610SAugustin Cavalier y[1] = (z-y[0]) - 4*pio2_1t;
109*f504f610SAugustin Cavalier return 4;
110*f504f610SAugustin Cavalier } else {
111*f504f610SAugustin Cavalier z = x + 4*pio2_1;
112*f504f610SAugustin Cavalier y[0] = z + 4*pio2_1t;
113*f504f610SAugustin Cavalier y[1] = (z-y[0]) + 4*pio2_1t;
114*f504f610SAugustin Cavalier return -4;
115*f504f610SAugustin Cavalier }
116*f504f610SAugustin Cavalier }
117*f504f610SAugustin Cavalier }
118*f504f610SAugustin Cavalier if (ix < 0x413921fb) { /* |x| ~< 2^20*(pi/2), medium size */
119*f504f610SAugustin Cavalier medium:
120*f504f610SAugustin Cavalier /* rint(x/(pi/2)), Assume round-to-nearest. */
121*f504f610SAugustin Cavalier fn = (double_t)x*invpio2 + toint - toint;
122*f504f610SAugustin Cavalier n = (int32_t)fn;
123*f504f610SAugustin Cavalier r = x - fn*pio2_1;
124*f504f610SAugustin Cavalier w = fn*pio2_1t; /* 1st round, good to 85 bits */
125*f504f610SAugustin Cavalier y[0] = r - w;
126*f504f610SAugustin Cavalier u.f = y[0];
127*f504f610SAugustin Cavalier ey = u.i>>52 & 0x7ff;
128*f504f610SAugustin Cavalier ex = ix>>20;
129*f504f610SAugustin Cavalier if (ex - ey > 16) { /* 2nd round, good to 118 bits */
130*f504f610SAugustin Cavalier t = r;
131*f504f610SAugustin Cavalier w = fn*pio2_2;
132*f504f610SAugustin Cavalier r = t - w;
133*f504f610SAugustin Cavalier w = fn*pio2_2t - ((t-r)-w);
134*f504f610SAugustin Cavalier y[0] = r - w;
135*f504f610SAugustin Cavalier u.f = y[0];
136*f504f610SAugustin Cavalier ey = u.i>>52 & 0x7ff;
137*f504f610SAugustin Cavalier if (ex - ey > 49) { /* 3rd round, good to 151 bits, covers all cases */
138*f504f610SAugustin Cavalier t = r;
139*f504f610SAugustin Cavalier w = fn*pio2_3;
140*f504f610SAugustin Cavalier r = t - w;
141*f504f610SAugustin Cavalier w = fn*pio2_3t - ((t-r)-w);
142*f504f610SAugustin Cavalier y[0] = r - w;
143*f504f610SAugustin Cavalier }
144*f504f610SAugustin Cavalier }
145*f504f610SAugustin Cavalier y[1] = (r - y[0]) - w;
146*f504f610SAugustin Cavalier return n;
147*f504f610SAugustin Cavalier }
148*f504f610SAugustin Cavalier /*
149*f504f610SAugustin Cavalier * all other (large) arguments
150*f504f610SAugustin Cavalier */
151*f504f610SAugustin Cavalier if (ix >= 0x7ff00000) { /* x is inf or NaN */
152*f504f610SAugustin Cavalier y[0] = y[1] = x - x;
153*f504f610SAugustin Cavalier return 0;
154*f504f610SAugustin Cavalier }
155*f504f610SAugustin Cavalier /* set z = scalbn(|x|,-ilogb(x)+23) */
156*f504f610SAugustin Cavalier u.f = x;
157*f504f610SAugustin Cavalier u.i &= (uint64_t)-1>>12;
158*f504f610SAugustin Cavalier u.i |= (uint64_t)(0x3ff + 23)<<52;
159*f504f610SAugustin Cavalier z = u.f;
160*f504f610SAugustin Cavalier for (i=0; i < 2; i++) {
161*f504f610SAugustin Cavalier tx[i] = (double)(int32_t)z;
162*f504f610SAugustin Cavalier z = (z-tx[i])*0x1p24;
163*f504f610SAugustin Cavalier }
164*f504f610SAugustin Cavalier tx[i] = z;
165*f504f610SAugustin Cavalier /* skip zero terms, first term is non-zero */
166*f504f610SAugustin Cavalier while (tx[i] == 0.0)
167*f504f610SAugustin Cavalier i--;
168*f504f610SAugustin Cavalier n = __rem_pio2_large(tx,ty,(int)(ix>>20)-(0x3ff+23),i+1,1);
169*f504f610SAugustin Cavalier if (sign) {
170*f504f610SAugustin Cavalier y[0] = -ty[0];
171*f504f610SAugustin Cavalier y[1] = -ty[1];
172*f504f610SAugustin Cavalier return -n;
173*f504f610SAugustin Cavalier }
174*f504f610SAugustin Cavalier y[0] = ty[0];
175*f504f610SAugustin Cavalier y[1] = ty[1];
176*f504f610SAugustin Cavalier return n;
177*f504f610SAugustin Cavalier }
178