1 | /* |
2 | * ==================================================== |
3 | * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved. |
4 | * |
5 | * Developed at SunPro, a Sun Microsystems, Inc. business. |
6 | * Permission to use, copy, modify, and distribute this |
7 | * software is freely granted, provided that this notice |
8 | * is preserved. |
9 | * ==================================================== |
10 | */ |
11 | |
12 | /* From e_hypotl.c -- long double version of e_hypot.c. |
13 | * Conversion to long double by Jakub Jelinek, jakub@redhat.com. |
14 | * Conversion to __float128 by FX Coudert, fxcoudert@gcc.gnu.org. |
15 | */ |
16 | |
17 | /* hypotq(x,y) |
18 | * |
19 | * Method : |
20 | * If (assume round-to-nearest) z=x*x+y*y |
21 | * has error less than sqrtl(2)/2 ulp, than |
22 | * sqrtl(z) has error less than 1 ulp (exercise). |
23 | * |
24 | * So, compute sqrtl(x*x+y*y) with some care as |
25 | * follows to get the error below 1 ulp: |
26 | * |
27 | * Assume x>y>0; |
28 | * (if possible, set rounding to round-to-nearest) |
29 | * 1. if x > 2y use |
30 | * x1*x1+(y*y+(x2*(x+x1))) for x*x+y*y |
31 | * where x1 = x with lower 64 bits cleared, x2 = x-x1; else |
32 | * 2. if x <= 2y use |
33 | * t1*y1+((x-y)*(x-y)+(t1*y2+t2*y)) |
34 | * where t1 = 2x with lower 64 bits cleared, t2 = 2x-t1, |
35 | * y1= y with lower 64 bits chopped, y2 = y-y1. |
36 | * |
37 | * NOTE: scaling may be necessary if some argument is too |
38 | * large or too tiny |
39 | * |
40 | * Special cases: |
41 | * hypotq(x,y) is INF if x or y is +INF or -INF; else |
42 | * hypotq(x,y) is NAN if x or y is NAN. |
43 | * |
44 | * Accuracy: |
45 | * hypotq(x,y) returns sqrtl(x^2+y^2) with error less |
46 | * than 1 ulps (units in the last place) |
47 | */ |
48 | |
49 | #include "quadmath-imp.h" |
50 | |
51 | __float128 |
52 | hypotq (__float128 x, __float128 y) |
53 | { |
54 | __float128 a, b, t1, t2, y1, y2, w; |
55 | int64_t j, k, ha, hb; |
56 | |
57 | GET_FLT128_MSW64(ha,x); |
58 | ha &= 0x7fffffffffffffffLL; |
59 | GET_FLT128_MSW64(hb,y); |
60 | hb &= 0x7fffffffffffffffLL; |
61 | if(hb > ha) {a=y;b=x;j=ha; ha=hb;hb=j;} else {a=x;b=y;} |
62 | SET_FLT128_MSW64(a,ha); /* a <- |a| */ |
63 | SET_FLT128_MSW64(b,hb); /* b <- |b| */ |
64 | if((ha-hb)>0x78000000000000LL) {return a+b;} /* x/y > 2**120 */ |
65 | k=0; |
66 | if(ha > 0x5f3f000000000000LL) { /* a>2**8000 */ |
67 | if(ha >= 0x7fff000000000000LL) { /* Inf or NaN */ |
68 | uint64_t low; |
69 | w = a+b; /* for sNaN */ |
70 | GET_FLT128_LSW64(low,a); |
71 | if(((ha&0xffffffffffffLL)|low)==0) w = a; |
72 | GET_FLT128_LSW64(low,b); |
73 | if(((hb^0x7fff000000000000LL)|low)==0) w = b; |
74 | return w; |
75 | } |
76 | /* scale a and b by 2**-9600 */ |
77 | ha -= 0x2580000000000000LL; |
78 | hb -= 0x2580000000000000LL; k += 9600; |
79 | SET_FLT128_MSW64(a,ha); |
80 | SET_FLT128_MSW64(b,hb); |
81 | } |
82 | if(hb < 0x20bf000000000000LL) { /* b < 2**-8000 */ |
83 | if(hb <= 0x0000ffffffffffffLL) { /* subnormal b or 0 */ |
84 | uint64_t low; |
85 | GET_FLT128_LSW64(low,b); |
86 | if((hb|low)==0) return a; |
87 | t1=0; |
88 | SET_FLT128_MSW64(t1,0x7ffd000000000000LL); /* t1=2^16382 */ |
89 | b *= t1; |
90 | a *= t1; |
91 | k -= 16382; |
92 | GET_FLT128_MSW64 (ha, a); |
93 | GET_FLT128_MSW64 (hb, b); |
94 | if (hb > ha) |
95 | { |
96 | t1 = a; |
97 | a = b; |
98 | b = t1; |
99 | j = ha; |
100 | ha = hb; |
101 | hb = j; |
102 | } |
103 | } else { /* scale a and b by 2^9600 */ |
104 | ha += 0x2580000000000000LL; /* a *= 2^9600 */ |
105 | hb += 0x2580000000000000LL; /* b *= 2^9600 */ |
106 | k -= 9600; |
107 | SET_FLT128_MSW64(a,ha); |
108 | SET_FLT128_MSW64(b,hb); |
109 | } |
110 | } |
111 | /* medium size a and b */ |
112 | w = a-b; |
113 | if (w>b) { |
114 | t1 = 0; |
115 | SET_FLT128_MSW64(t1,ha); |
116 | t2 = a-t1; |
117 | w = sqrtq(t1*t1-(b*(-b)-t2*(a+t1))); |
118 | } else { |
119 | a = a+a; |
120 | y1 = 0; |
121 | SET_FLT128_MSW64(y1,hb); |
122 | y2 = b - y1; |
123 | t1 = 0; |
124 | SET_FLT128_MSW64(t1,ha+0x0001000000000000LL); |
125 | t2 = a - t1; |
126 | w = sqrtq(t1*y1-(w*(-w)-(t1*y2+t2*b))); |
127 | } |
128 | if(k!=0) { |
129 | uint64_t high; |
130 | t1 = 1.0Q; |
131 | GET_FLT128_MSW64(high,t1); |
132 | SET_FLT128_MSW64(t1,high+(k<<48)); |
133 | w *= t1; |
134 | math_check_force_underflow_nonneg (w); |
135 | return w; |
136 | } else return w; |
137 | } |
138 | |