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