xref: /openbmc/linux/arch/mips/math-emu/sp_maddf.c (revision c4c3c32d)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * IEEE754 floating point arithmetic
4  * single precision: MADDF.f (Fused Multiply Add)
5  * MADDF.fmt: FPR[fd] = FPR[fd] + (FPR[fs] x FPR[ft])
6  *
7  * MIPS floating point support
8  * Copyright (C) 2015 Imagination Technologies, Ltd.
9  * Author: Markos Chandras <markos.chandras@imgtec.com>
10  */
11 
12 #include "ieee754sp.h"
13 
14 
15 static union ieee754sp _sp_maddf(union ieee754sp z, union ieee754sp x,
16 				 union ieee754sp y, enum maddf_flags flags)
17 {
18 	int re;
19 	int rs;
20 	unsigned int rm;
21 	u64 rm64;
22 	u64 zm64;
23 	int s;
24 
25 	COMPXSP;
26 	COMPYSP;
27 	COMPZSP;
28 
29 	EXPLODEXSP;
30 	EXPLODEYSP;
31 	EXPLODEZSP;
32 
33 	FLUSHXSP;
34 	FLUSHYSP;
35 	FLUSHZSP;
36 
37 	ieee754_clearcx();
38 
39 	rs = xs ^ ys;
40 	if (flags & MADDF_NEGATE_PRODUCT)
41 		rs ^= 1;
42 	if (flags & MADDF_NEGATE_ADDITION)
43 		zs ^= 1;
44 
45 	/*
46 	 * Handle the cases when at least one of x, y or z is a NaN.
47 	 * Order of precedence is sNaN, qNaN and z, x, y.
48 	 */
49 	if (zc == IEEE754_CLASS_SNAN)
50 		return ieee754sp_nanxcpt(z);
51 	if (xc == IEEE754_CLASS_SNAN)
52 		return ieee754sp_nanxcpt(x);
53 	if (yc == IEEE754_CLASS_SNAN)
54 		return ieee754sp_nanxcpt(y);
55 	if (zc == IEEE754_CLASS_QNAN)
56 		return z;
57 	if (xc == IEEE754_CLASS_QNAN)
58 		return x;
59 	if (yc == IEEE754_CLASS_QNAN)
60 		return y;
61 
62 	if (zc == IEEE754_CLASS_DNORM)
63 		SPDNORMZ;
64 	/* ZERO z cases are handled separately below */
65 
66 	switch (CLPAIR(xc, yc)) {
67 
68 
69 	/*
70 	 * Infinity handling
71 	 */
72 	case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_ZERO):
73 	case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_INF):
74 		ieee754_setcx(IEEE754_INVALID_OPERATION);
75 		return ieee754sp_indef();
76 
77 	case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_INF):
78 	case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_INF):
79 	case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_NORM):
80 	case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_DNORM):
81 	case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_INF):
82 		if ((zc == IEEE754_CLASS_INF) && (zs != rs)) {
83 			/*
84 			 * Cases of addition of infinities with opposite signs
85 			 * or subtraction of infinities with same signs.
86 			 */
87 			ieee754_setcx(IEEE754_INVALID_OPERATION);
88 			return ieee754sp_indef();
89 		}
90 		/*
91 		 * z is here either not an infinity, or an infinity having the
92 		 * same sign as product (x*y). The result must be an infinity,
93 		 * and its sign is determined only by the sign of product (x*y).
94 		 */
95 		return ieee754sp_inf(rs);
96 
97 	case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_ZERO):
98 	case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_NORM):
99 	case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_DNORM):
100 	case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_ZERO):
101 	case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_ZERO):
102 		if (zc == IEEE754_CLASS_INF)
103 			return ieee754sp_inf(zs);
104 		if (zc == IEEE754_CLASS_ZERO) {
105 			/* Handle cases +0 + (-0) and similar ones. */
106 			if (zs == rs)
107 				/*
108 				 * Cases of addition of zeros of equal signs
109 				 * or subtraction of zeroes of opposite signs.
110 				 * The sign of the resulting zero is in any
111 				 * such case determined only by the sign of z.
112 				 */
113 				return z;
114 
115 			return ieee754sp_zero(ieee754_csr.rm == FPU_CSR_RD);
116 		}
117 		/* x*y is here 0, and z is not 0, so just return z */
118 		return z;
119 
120 	case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_DNORM):
121 		SPDNORMX;
122 		fallthrough;
123 	case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_DNORM):
124 		if (zc == IEEE754_CLASS_INF)
125 			return ieee754sp_inf(zs);
126 		SPDNORMY;
127 		break;
128 
129 	case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_NORM):
130 		if (zc == IEEE754_CLASS_INF)
131 			return ieee754sp_inf(zs);
132 		SPDNORMX;
133 		break;
134 
135 	case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_NORM):
136 		if (zc == IEEE754_CLASS_INF)
137 			return ieee754sp_inf(zs);
138 		/* continue to real computations */
139 	}
140 
141 	/* Finally get to do some computation */
142 
143 	/*
144 	 * Do the multiplication bit first
145 	 *
146 	 * rm = xm * ym, re = xe + ye basically
147 	 *
148 	 * At this point xm and ym should have been normalized.
149 	 */
150 
151 	/* rm = xm * ym, re = xe+ye basically */
152 	assert(xm & SP_HIDDEN_BIT);
153 	assert(ym & SP_HIDDEN_BIT);
154 
155 	re = xe + ye;
156 
157 	/* Multiple 24 bit xm and ym to give 48 bit results */
158 	rm64 = (uint64_t)xm * ym;
159 
160 	/* Shunt to top of word */
161 	rm64 = rm64 << 16;
162 
163 	/* Put explicit bit at bit 62 if necessary */
164 	if ((int64_t) rm64 < 0) {
165 		rm64 = rm64 >> 1;
166 		re++;
167 	}
168 
169 	assert(rm64 & (1 << 62));
170 
171 	if (zc == IEEE754_CLASS_ZERO) {
172 		/*
173 		 * Move explicit bit from bit 62 to bit 26 since the
174 		 * ieee754sp_format code expects the mantissa to be
175 		 * 27 bits wide (24 + 3 rounding bits).
176 		 */
177 		rm = XSPSRS64(rm64, (62 - 26));
178 		return ieee754sp_format(rs, re, rm);
179 	}
180 
181 	/* Move explicit bit from bit 23 to bit 62 */
182 	zm64 = (uint64_t)zm << (62 - 23);
183 	assert(zm64 & (1 << 62));
184 
185 	/* Make the exponents the same */
186 	if (ze > re) {
187 		/*
188 		 * Have to shift r fraction right to align.
189 		 */
190 		s = ze - re;
191 		rm64 = XSPSRS64(rm64, s);
192 		re += s;
193 	} else if (re > ze) {
194 		/*
195 		 * Have to shift z fraction right to align.
196 		 */
197 		s = re - ze;
198 		zm64 = XSPSRS64(zm64, s);
199 		ze += s;
200 	}
201 	assert(ze == re);
202 	assert(ze <= SP_EMAX);
203 
204 	/* Do the addition */
205 	if (zs == rs) {
206 		/*
207 		 * Generate 64 bit result by adding two 63 bit numbers
208 		 * leaving result in zm64, zs and ze.
209 		 */
210 		zm64 = zm64 + rm64;
211 		if ((int64_t)zm64 < 0) {	/* carry out */
212 			zm64 = XSPSRS1(zm64);
213 			ze++;
214 		}
215 	} else {
216 		if (zm64 >= rm64) {
217 			zm64 = zm64 - rm64;
218 		} else {
219 			zm64 = rm64 - zm64;
220 			zs = rs;
221 		}
222 		if (zm64 == 0)
223 			return ieee754sp_zero(ieee754_csr.rm == FPU_CSR_RD);
224 
225 		/*
226 		 * Put explicit bit at bit 62 if necessary.
227 		 */
228 		while ((zm64 >> 62) == 0) {
229 			zm64 <<= 1;
230 			ze--;
231 		}
232 	}
233 
234 	/*
235 	 * Move explicit bit from bit 62 to bit 26 since the
236 	 * ieee754sp_format code expects the mantissa to be
237 	 * 27 bits wide (24 + 3 rounding bits).
238 	 */
239 	zm = XSPSRS64(zm64, (62 - 26));
240 
241 	return ieee754sp_format(zs, ze, zm);
242 }
243 
244 union ieee754sp ieee754sp_maddf(union ieee754sp z, union ieee754sp x,
245 				union ieee754sp y)
246 {
247 	return _sp_maddf(z, x, y, 0);
248 }
249 
250 union ieee754sp ieee754sp_msubf(union ieee754sp z, union ieee754sp x,
251 				union ieee754sp y)
252 {
253 	return _sp_maddf(z, x, y, MADDF_NEGATE_PRODUCT);
254 }
255 
256 union ieee754sp ieee754sp_madd(union ieee754sp z, union ieee754sp x,
257 				union ieee754sp y)
258 {
259 	return _sp_maddf(z, x, y, 0);
260 }
261 
262 union ieee754sp ieee754sp_msub(union ieee754sp z, union ieee754sp x,
263 				union ieee754sp y)
264 {
265 	return _sp_maddf(z, x, y, MADDF_NEGATE_ADDITION);
266 }
267 
268 union ieee754sp ieee754sp_nmadd(union ieee754sp z, union ieee754sp x,
269 				union ieee754sp y)
270 {
271 	return _sp_maddf(z, x, y, MADDF_NEGATE_PRODUCT|MADDF_NEGATE_ADDITION);
272 }
273 
274 union ieee754sp ieee754sp_nmsub(union ieee754sp z, union ieee754sp x,
275 				union ieee754sp y)
276 {
277 	return _sp_maddf(z, x, y, MADDF_NEGATE_PRODUCT);
278 }
279