fpu_add.c revision 1.7 1 1.7 rin /* $NetBSD: fpu_add.c,v 1.7 2022/08/28 22:09:26 rin Exp $ */
2 1.2 deraadt
3 1.1 deraadt /*
4 1.1 deraadt * Copyright (c) 1992, 1993
5 1.1 deraadt * The Regents of the University of California. All rights reserved.
6 1.1 deraadt *
7 1.1 deraadt * This software was developed by the Computer Systems Engineering group
8 1.1 deraadt * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and
9 1.1 deraadt * contributed to Berkeley.
10 1.1 deraadt *
11 1.1 deraadt * All advertising materials mentioning features or use of this software
12 1.1 deraadt * must display the following acknowledgement:
13 1.1 deraadt * This product includes software developed by the University of
14 1.1 deraadt * California, Lawrence Berkeley Laboratory.
15 1.1 deraadt *
16 1.1 deraadt * Redistribution and use in source and binary forms, with or without
17 1.1 deraadt * modification, are permitted provided that the following conditions
18 1.1 deraadt * are met:
19 1.1 deraadt * 1. Redistributions of source code must retain the above copyright
20 1.1 deraadt * notice, this list of conditions and the following disclaimer.
21 1.1 deraadt * 2. Redistributions in binary form must reproduce the above copyright
22 1.1 deraadt * notice, this list of conditions and the following disclaimer in the
23 1.1 deraadt * documentation and/or other materials provided with the distribution.
24 1.5 agc * 3. Neither the name of the University nor the names of its contributors
25 1.1 deraadt * may be used to endorse or promote products derived from this software
26 1.1 deraadt * without specific prior written permission.
27 1.1 deraadt *
28 1.1 deraadt * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29 1.1 deraadt * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30 1.1 deraadt * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31 1.1 deraadt * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32 1.1 deraadt * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33 1.1 deraadt * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34 1.1 deraadt * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35 1.1 deraadt * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36 1.1 deraadt * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37 1.1 deraadt * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38 1.1 deraadt * SUCH DAMAGE.
39 1.1 deraadt *
40 1.1 deraadt * @(#)fpu_add.c 8.1 (Berkeley) 6/11/93
41 1.1 deraadt */
42 1.1 deraadt
43 1.1 deraadt /*
44 1.1 deraadt * Perform an FPU add (return x + y).
45 1.1 deraadt *
46 1.1 deraadt * To subtract, negate y and call add.
47 1.1 deraadt */
48 1.4 lukem
49 1.4 lukem #include <sys/cdefs.h>
50 1.7 rin __KERNEL_RCSID(0, "$NetBSD: fpu_add.c,v 1.7 2022/08/28 22:09:26 rin Exp $");
51 1.1 deraadt
52 1.1 deraadt #include <sys/types.h>
53 1.3 christos #ifdef DIAGNOSTIC
54 1.3 christos #include <sys/systm.h>
55 1.3 christos #endif
56 1.1 deraadt
57 1.1 deraadt #include <machine/reg.h>
58 1.3 christos #include <machine/instr.h>
59 1.1 deraadt
60 1.1 deraadt #include <sparc/fpu/fpu_arith.h>
61 1.1 deraadt #include <sparc/fpu/fpu_emu.h>
62 1.3 christos #include <sparc/fpu/fpu_extern.h>
63 1.1 deraadt
64 1.1 deraadt struct fpn *
65 1.6 uwe fpu_add(struct fpemu *fe)
66 1.1 deraadt {
67 1.7 rin struct fpn *x = &fe->fe_f1, *y = &fe->fe_f2, *r;
68 1.7 rin u_int r0, r1, r2, r3;
69 1.7 rin int rd;
70 1.1 deraadt
71 1.1 deraadt /*
72 1.1 deraadt * Put the `heavier' operand on the right (see fpu_emu.h).
73 1.1 deraadt * Then we will have one of the following cases, taken in the
74 1.1 deraadt * following order:
75 1.1 deraadt *
76 1.1 deraadt * - y = NaN. Implied: if only one is a signalling NaN, y is.
77 1.1 deraadt * The result is y.
78 1.1 deraadt * - y = Inf. Implied: x != NaN (is 0, number, or Inf: the NaN
79 1.1 deraadt * case was taken care of earlier).
80 1.1 deraadt * If x = -y, the result is NaN. Otherwise the result
81 1.1 deraadt * is y (an Inf of whichever sign).
82 1.1 deraadt * - y is 0. Implied: x = 0.
83 1.1 deraadt * If x and y differ in sign (one positive, one negative),
84 1.1 deraadt * the result is +0 except when rounding to -Inf. If same:
85 1.1 deraadt * +0 + +0 = +0; -0 + -0 = -0.
86 1.1 deraadt * - x is 0. Implied: y != 0.
87 1.1 deraadt * Result is y.
88 1.1 deraadt * - other. Implied: both x and y are numbers.
89 1.1 deraadt * Do addition a la Hennessey & Patterson.
90 1.1 deraadt */
91 1.1 deraadt ORDER(x, y);
92 1.1 deraadt if (ISNAN(y))
93 1.1 deraadt return (y);
94 1.1 deraadt if (ISINF(y)) {
95 1.1 deraadt if (ISINF(x) && x->fp_sign != y->fp_sign)
96 1.1 deraadt return (fpu_newnan(fe));
97 1.1 deraadt return (y);
98 1.1 deraadt }
99 1.1 deraadt rd = ((fe->fe_fsr >> FSR_RD_SHIFT) & FSR_RD_MASK);
100 1.1 deraadt if (ISZERO(y)) {
101 1.1 deraadt if (rd != FSR_RD_RM) /* only -0 + -0 gives -0 */
102 1.1 deraadt y->fp_sign &= x->fp_sign;
103 1.1 deraadt else /* any -0 operand gives -0 */
104 1.1 deraadt y->fp_sign |= x->fp_sign;
105 1.1 deraadt return (y);
106 1.1 deraadt }
107 1.1 deraadt if (ISZERO(x))
108 1.1 deraadt return (y);
109 1.1 deraadt /*
110 1.1 deraadt * We really have two numbers to add, although their signs may
111 1.1 deraadt * differ. Make the exponents match, by shifting the smaller
112 1.1 deraadt * number right (e.g., 1.011 => 0.1011) and increasing its
113 1.1 deraadt * exponent (2^3 => 2^4). Note that we do not alter the exponents
114 1.1 deraadt * of x and y here.
115 1.1 deraadt */
116 1.1 deraadt r = &fe->fe_f3;
117 1.1 deraadt r->fp_class = FPC_NUM;
118 1.1 deraadt if (x->fp_exp == y->fp_exp) {
119 1.1 deraadt r->fp_exp = x->fp_exp;
120 1.1 deraadt r->fp_sticky = 0;
121 1.1 deraadt } else {
122 1.1 deraadt if (x->fp_exp < y->fp_exp) {
123 1.1 deraadt /*
124 1.1 deraadt * Try to avoid subtract case iii (see below).
125 1.1 deraadt * This also guarantees that x->fp_sticky = 0.
126 1.1 deraadt */
127 1.1 deraadt SWAP(x, y);
128 1.1 deraadt }
129 1.1 deraadt /* now x->fp_exp > y->fp_exp */
130 1.1 deraadt r->fp_exp = x->fp_exp;
131 1.1 deraadt r->fp_sticky = fpu_shr(y, x->fp_exp - y->fp_exp);
132 1.1 deraadt }
133 1.1 deraadt r->fp_sign = x->fp_sign;
134 1.1 deraadt if (x->fp_sign == y->fp_sign) {
135 1.1 deraadt FPU_DECL_CARRY
136 1.1 deraadt
137 1.1 deraadt /*
138 1.1 deraadt * The signs match, so we simply add the numbers. The result
139 1.1 deraadt * may be `supernormal' (as big as 1.111...1 + 1.111...1, or
140 1.1 deraadt * 11.111...0). If so, a single bit shift-right will fix it
141 1.1 deraadt * (but remember to adjust the exponent).
142 1.1 deraadt */
143 1.1 deraadt /* r->fp_mant = x->fp_mant + y->fp_mant */
144 1.1 deraadt FPU_ADDS(r->fp_mant[3], x->fp_mant[3], y->fp_mant[3]);
145 1.1 deraadt FPU_ADDCS(r->fp_mant[2], x->fp_mant[2], y->fp_mant[2]);
146 1.1 deraadt FPU_ADDCS(r->fp_mant[1], x->fp_mant[1], y->fp_mant[1]);
147 1.1 deraadt FPU_ADDC(r0, x->fp_mant[0], y->fp_mant[0]);
148 1.1 deraadt if ((r->fp_mant[0] = r0) >= FP_2) {
149 1.1 deraadt (void) fpu_shr(r, 1);
150 1.1 deraadt r->fp_exp++;
151 1.1 deraadt }
152 1.1 deraadt } else {
153 1.1 deraadt FPU_DECL_CARRY
154 1.1 deraadt
155 1.1 deraadt /*
156 1.1 deraadt * The signs differ, so things are rather more difficult.
157 1.1 deraadt * H&P would have us negate the negative operand and add;
158 1.1 deraadt * this is the same as subtracting the negative operand.
159 1.1 deraadt * This is quite a headache. Instead, we will subtract
160 1.1 deraadt * y from x, regardless of whether y itself is the negative
161 1.1 deraadt * operand. When this is done one of three conditions will
162 1.1 deraadt * hold, depending on the magnitudes of x and y:
163 1.1 deraadt * case i) |x| > |y|. The result is just x - y,
164 1.1 deraadt * with x's sign, but it may need to be normalized.
165 1.1 deraadt * case ii) |x| = |y|. The result is 0 (maybe -0)
166 1.1 deraadt * so must be fixed up.
167 1.1 deraadt * case iii) |x| < |y|. We goofed; the result should
168 1.1 deraadt * be (y - x), with the same sign as y.
169 1.1 deraadt * We could compare |x| and |y| here and avoid case iii,
170 1.1 deraadt * but that would take just as much work as the subtract.
171 1.1 deraadt * We can tell case iii has occurred by an overflow.
172 1.1 deraadt *
173 1.1 deraadt * N.B.: since x->fp_exp >= y->fp_exp, x->fp_sticky = 0.
174 1.1 deraadt */
175 1.1 deraadt /* r->fp_mant = x->fp_mant - y->fp_mant */
176 1.1 deraadt FPU_SET_CARRY(y->fp_sticky);
177 1.1 deraadt FPU_SUBCS(r3, x->fp_mant[3], y->fp_mant[3]);
178 1.1 deraadt FPU_SUBCS(r2, x->fp_mant[2], y->fp_mant[2]);
179 1.1 deraadt FPU_SUBCS(r1, x->fp_mant[1], y->fp_mant[1]);
180 1.1 deraadt FPU_SUBC(r0, x->fp_mant[0], y->fp_mant[0]);
181 1.1 deraadt if (r0 < FP_2) {
182 1.1 deraadt /* cases i and ii */
183 1.1 deraadt if ((r0 | r1 | r2 | r3) == 0) {
184 1.1 deraadt /* case ii */
185 1.1 deraadt r->fp_class = FPC_ZERO;
186 1.1 deraadt r->fp_sign = rd == FSR_RD_RM;
187 1.1 deraadt return (r);
188 1.1 deraadt }
189 1.1 deraadt } else {
190 1.1 deraadt /*
191 1.1 deraadt * Oops, case iii. This can only occur when the
192 1.1 deraadt * exponents were equal, in which case neither
193 1.1 deraadt * x nor y have sticky bits set. Flip the sign
194 1.1 deraadt * (to y's sign) and negate the result to get y - x.
195 1.1 deraadt */
196 1.1 deraadt #ifdef DIAGNOSTIC
197 1.1 deraadt if (x->fp_exp != y->fp_exp || r->fp_sticky)
198 1.1 deraadt panic("fpu_add");
199 1.1 deraadt #endif
200 1.1 deraadt r->fp_sign = y->fp_sign;
201 1.1 deraadt FPU_SUBS(r3, 0, r3);
202 1.1 deraadt FPU_SUBCS(r2, 0, r2);
203 1.1 deraadt FPU_SUBCS(r1, 0, r1);
204 1.1 deraadt FPU_SUBC(r0, 0, r0);
205 1.1 deraadt }
206 1.1 deraadt r->fp_mant[3] = r3;
207 1.1 deraadt r->fp_mant[2] = r2;
208 1.1 deraadt r->fp_mant[1] = r1;
209 1.1 deraadt r->fp_mant[0] = r0;
210 1.1 deraadt if (r0 < FP_1)
211 1.1 deraadt fpu_norm(r);
212 1.1 deraadt }
213 1.1 deraadt return (r);
214 1.1 deraadt }
215