1*627f7eb2Smrg /* e_hypotl.c -- long double version of e_hypot.c.
2*627f7eb2Smrg * Conversion to long double by Jakub Jelinek, jakub@redhat.com.
3*627f7eb2Smrg */
4*627f7eb2Smrg
5*627f7eb2Smrg /*
6*627f7eb2Smrg * ====================================================
7*627f7eb2Smrg * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
8*627f7eb2Smrg *
9*627f7eb2Smrg * Developed at SunPro, a Sun Microsystems, Inc. business.
10*627f7eb2Smrg * Permission to use, copy, modify, and distribute this
11*627f7eb2Smrg * software is freely granted, provided that this notice
12*627f7eb2Smrg * is preserved.
13*627f7eb2Smrg * ====================================================
14*627f7eb2Smrg */
15*627f7eb2Smrg
16*627f7eb2Smrg /* hypotq(x,y)
17*627f7eb2Smrg *
18*627f7eb2Smrg * Method :
19*627f7eb2Smrg * If (assume round-to-nearest) z=x*x+y*y
20*627f7eb2Smrg * has error less than sqrtq(2)/2 ulp, than
21*627f7eb2Smrg * sqrtq(z) has error less than 1 ulp (exercise).
22*627f7eb2Smrg *
23*627f7eb2Smrg * So, compute sqrtq(x*x+y*y) with some care as
24*627f7eb2Smrg * follows to get the error below 1 ulp:
25*627f7eb2Smrg *
26*627f7eb2Smrg * Assume x>y>0;
27*627f7eb2Smrg * (if possible, set rounding to round-to-nearest)
28*627f7eb2Smrg * 1. if x > 2y use
29*627f7eb2Smrg * x1*x1+(y*y+(x2*(x+x1))) for x*x+y*y
30*627f7eb2Smrg * where x1 = x with lower 64 bits cleared, x2 = x-x1; else
31*627f7eb2Smrg * 2. if x <= 2y use
32*627f7eb2Smrg * t1*y1+((x-y)*(x-y)+(t1*y2+t2*y))
33*627f7eb2Smrg * where t1 = 2x with lower 64 bits cleared, t2 = 2x-t1,
34*627f7eb2Smrg * y1= y with lower 64 bits chopped, y2 = y-y1.
35*627f7eb2Smrg *
36*627f7eb2Smrg * NOTE: scaling may be necessary if some argument is too
37*627f7eb2Smrg * large or too tiny
38*627f7eb2Smrg *
39*627f7eb2Smrg * Special cases:
40*627f7eb2Smrg * hypotl(x,y) is INF if x or y is +INF or -INF; else
41*627f7eb2Smrg * hypotl(x,y) is NAN if x or y is NAN.
42*627f7eb2Smrg *
43*627f7eb2Smrg * Accuracy:
44*627f7eb2Smrg * hypotl(x,y) returns sqrtq(x^2+y^2) with error less
45*627f7eb2Smrg * than 1 ulps (units in the last place)
46*627f7eb2Smrg */
47*627f7eb2Smrg
48*627f7eb2Smrg #include "quadmath-imp.h"
49*627f7eb2Smrg
50*627f7eb2Smrg __float128
hypotq(__float128 x,__float128 y)51*627f7eb2Smrg hypotq(__float128 x, __float128 y)
52*627f7eb2Smrg {
53*627f7eb2Smrg __float128 a,b,t1,t2,y1,y2,w;
54*627f7eb2Smrg int64_t j,k,ha,hb;
55*627f7eb2Smrg
56*627f7eb2Smrg GET_FLT128_MSW64(ha,x);
57*627f7eb2Smrg ha &= 0x7fffffffffffffffLL;
58*627f7eb2Smrg GET_FLT128_MSW64(hb,y);
59*627f7eb2Smrg hb &= 0x7fffffffffffffffLL;
60*627f7eb2Smrg if(hb > ha) {a=y;b=x;j=ha; ha=hb;hb=j;} else {a=x;b=y;}
61*627f7eb2Smrg SET_FLT128_MSW64(a,ha); /* a <- |a| */
62*627f7eb2Smrg SET_FLT128_MSW64(b,hb); /* b <- |b| */
63*627f7eb2Smrg if((ha-hb)>0x78000000000000LL) {return a+b;} /* x/y > 2**120 */
64*627f7eb2Smrg k=0;
65*627f7eb2Smrg if(ha > 0x5f3f000000000000LL) { /* a>2**8000 */
66*627f7eb2Smrg if(ha >= 0x7fff000000000000LL) { /* Inf or NaN */
67*627f7eb2Smrg uint64_t low;
68*627f7eb2Smrg w = a+b; /* for sNaN */
69*627f7eb2Smrg if (issignalingq (a) || issignalingq (b))
70*627f7eb2Smrg return w;
71*627f7eb2Smrg GET_FLT128_LSW64(low,a);
72*627f7eb2Smrg if(((ha&0xffffffffffffLL)|low)==0) w = a;
73*627f7eb2Smrg GET_FLT128_LSW64(low,b);
74*627f7eb2Smrg if(((hb^0x7fff000000000000LL)|low)==0) w = b;
75*627f7eb2Smrg return w;
76*627f7eb2Smrg }
77*627f7eb2Smrg /* scale a and b by 2**-9600 */
78*627f7eb2Smrg ha -= 0x2580000000000000LL;
79*627f7eb2Smrg hb -= 0x2580000000000000LL; k += 9600;
80*627f7eb2Smrg SET_FLT128_MSW64(a,ha);
81*627f7eb2Smrg SET_FLT128_MSW64(b,hb);
82*627f7eb2Smrg }
83*627f7eb2Smrg if(hb < 0x20bf000000000000LL) { /* b < 2**-8000 */
84*627f7eb2Smrg if(hb <= 0x0000ffffffffffffLL) { /* subnormal b or 0 */
85*627f7eb2Smrg uint64_t low;
86*627f7eb2Smrg GET_FLT128_LSW64(low,b);
87*627f7eb2Smrg if((hb|low)==0) return a;
88*627f7eb2Smrg t1=0;
89*627f7eb2Smrg SET_FLT128_MSW64(t1,0x7ffd000000000000LL); /* t1=2^16382 */
90*627f7eb2Smrg b *= t1;
91*627f7eb2Smrg a *= t1;
92*627f7eb2Smrg k -= 16382;
93*627f7eb2Smrg GET_FLT128_MSW64 (ha, a);
94*627f7eb2Smrg GET_FLT128_MSW64 (hb, b);
95*627f7eb2Smrg if (hb > ha)
96*627f7eb2Smrg {
97*627f7eb2Smrg t1 = a;
98*627f7eb2Smrg a = b;
99*627f7eb2Smrg b = t1;
100*627f7eb2Smrg j = ha;
101*627f7eb2Smrg ha = hb;
102*627f7eb2Smrg hb = j;
103*627f7eb2Smrg }
104*627f7eb2Smrg } else { /* scale a and b by 2^9600 */
105*627f7eb2Smrg ha += 0x2580000000000000LL; /* a *= 2^9600 */
106*627f7eb2Smrg hb += 0x2580000000000000LL; /* b *= 2^9600 */
107*627f7eb2Smrg k -= 9600;
108*627f7eb2Smrg SET_FLT128_MSW64(a,ha);
109*627f7eb2Smrg SET_FLT128_MSW64(b,hb);
110*627f7eb2Smrg }
111*627f7eb2Smrg }
112*627f7eb2Smrg /* medium size a and b */
113*627f7eb2Smrg w = a-b;
114*627f7eb2Smrg if (w>b) {
115*627f7eb2Smrg t1 = 0;
116*627f7eb2Smrg SET_FLT128_MSW64(t1,ha);
117*627f7eb2Smrg t2 = a-t1;
118*627f7eb2Smrg w = sqrtq(t1*t1-(b*(-b)-t2*(a+t1)));
119*627f7eb2Smrg } else {
120*627f7eb2Smrg a = a+a;
121*627f7eb2Smrg y1 = 0;
122*627f7eb2Smrg SET_FLT128_MSW64(y1,hb);
123*627f7eb2Smrg y2 = b - y1;
124*627f7eb2Smrg t1 = 0;
125*627f7eb2Smrg SET_FLT128_MSW64(t1,ha+0x0001000000000000LL);
126*627f7eb2Smrg t2 = a - t1;
127*627f7eb2Smrg w = sqrtq(t1*y1-(w*(-w)-(t1*y2+t2*b)));
128*627f7eb2Smrg }
129*627f7eb2Smrg if(k!=0) {
130*627f7eb2Smrg uint64_t high;
131*627f7eb2Smrg t1 = 1;
132*627f7eb2Smrg GET_FLT128_MSW64(high,t1);
133*627f7eb2Smrg SET_FLT128_MSW64(t1,high+(k<<48));
134*627f7eb2Smrg w *= t1;
135*627f7eb2Smrg math_check_force_underflow_nonneg (w);
136*627f7eb2Smrg return w;
137*627f7eb2Smrg } else return w;
138*627f7eb2Smrg }
139