1 // SPDX-License-Identifier: GPL-2.0-only
2 /* IEEE754 floating point arithmetic
3 * double precision square root
6 * MIPS floating point support
7 * Copyright (C) 1994-2000 Algorithmics Ltd.
10 #include "ieee754dp.h"
12 static const unsigned int table[] = {
13 0, 1204, 3062, 5746, 9193, 13348, 18162, 23592,
14 29598, 36145, 43202, 50740, 58733, 67158, 75992,
15 85215, 83599, 71378, 60428, 50647, 41945, 34246,
16 27478, 21581, 16499, 12183, 8588, 5674, 3403,
20 union ieee754dp ieee754dp_sqrt(union ieee754dp x)
22 struct _ieee754_csr oldcsr;
23 union ieee754dp y, z, t;
24 unsigned int scalx, yh;
31 /* x == INF or NAN? */
33 case IEEE754_CLASS_SNAN:
34 return ieee754dp_nanxcpt(x);
36 case IEEE754_CLASS_QNAN:
40 case IEEE754_CLASS_ZERO:
44 case IEEE754_CLASS_INF:
46 /* sqrt(-Inf) = Nan */
47 ieee754_setcx(IEEE754_INVALID_OPERATION);
48 return ieee754dp_indef();
50 /* sqrt(+Inf) = Inf */
53 case IEEE754_CLASS_DNORM:
56 case IEEE754_CLASS_NORM:
59 ieee754_setcx(IEEE754_INVALID_OPERATION);
60 return ieee754dp_indef();
65 /* save old csr; switch off INX enable & flag; set RN rounding */
67 ieee754_csr.mx &= ~IEEE754_INEXACT;
68 ieee754_csr.sx &= ~IEEE754_INEXACT;
69 ieee754_csr.rm = FPU_CSR_RN;
71 /* adjust exponent to prevent overflow */
73 if (xe > 512) { /* x > 2**-512? */
74 xe -= 512; /* x = x / 2**512 */
76 } else if (xe < -512) { /* x < 2**-512? */
77 xe += 512; /* x = x * 2**512 */
81 x = builddp(0, xe + DP_EBIAS, xm & ~DP_HIDDEN_BIT);
84 /* magic initial approximation to almost 8 sig. bits */
86 yh = (yh >> 1) + 0x1ff80000;
87 yh = yh - table[(yh >> 15) & 31];
88 y.bits = ((u64) yh << 32) | (y.bits & 0xffffffff);
90 /* Heron's rule once with correction to improve to ~18 sig. bits */
91 /* t=x/y; y=y+t; py[n0]=py[n0]-0x00100006; py[n1]=0; */
92 t = ieee754dp_div(x, y);
93 y = ieee754dp_add(y, t);
94 y.bits -= 0x0010000600000000LL;
95 y.bits &= 0xffffffff00000000LL;
97 /* triple to almost 56 sig. bits: y ~= sqrt(x) to within 1 ulp */
98 /* t=y*y; z=t; pt[n0]+=0x00100000; t+=z; z=(x-z)*y; */
99 t = ieee754dp_mul(y, y);
102 t = ieee754dp_add(t, z);
103 z = ieee754dp_mul(ieee754dp_sub(x, z), y);
105 /* t=z/(t+x) ; pt[n0]+=0x00100000; y+=t; */
106 t = ieee754dp_div(z, ieee754dp_add(t, x));
108 y = ieee754dp_add(y, t);
110 /* twiddle last bit to force y correctly rounded */
112 /* set RZ, clear INEX flag */
113 ieee754_csr.rm = FPU_CSR_RZ;
114 ieee754_csr.sx &= ~IEEE754_INEXACT;
116 /* t=x/y; ...chopped quotient, possibly inexact */
117 t = ieee754dp_div(x, y);
119 if (ieee754_csr.sx & IEEE754_INEXACT || t.bits != y.bits) {
121 if (!(ieee754_csr.sx & IEEE754_INEXACT))
125 /* add inexact to result status */
126 oldcsr.cx |= IEEE754_INEXACT;
127 oldcsr.sx |= IEEE754_INEXACT;
138 /* y=y+t; ...chopped sum */
139 y = ieee754dp_add(y, t);
141 /* adjust scalx for correctly rounded sqrt(x) */
145 /* py[n0]=py[n0]+scalx; ...scale back y */
148 /* restore rounding mode, possibly set inexact */
149 ieee754_csr = oldcsr;