xref: /openbmc/linux/arch/mips/math-emu/sp_maddf.c (revision d4fd6347)
1 /*
2  * IEEE754 floating point arithmetic
3  * single precision: MADDF.f (Fused Multiply Add)
4  * MADDF.fmt: FPR[fd] = FPR[fd] + (FPR[fs] x FPR[ft])
5  *
6  * MIPS floating point support
7  * Copyright (C) 2015 Imagination Technologies, Ltd.
8  * Author: Markos Chandras <markos.chandras@imgtec.com>
9  *
10  *  This program is free software; you can distribute it and/or modify it
11  *  under the terms of the GNU General Public License as published by the
12  *  Free Software Foundation; version 2 of the License.
13  */
14 
15 #include "ieee754sp.h"
16 
17 
18 static union ieee754sp _sp_maddf(union ieee754sp z, union ieee754sp x,
19 				 union ieee754sp y, enum maddf_flags flags)
20 {
21 	int re;
22 	int rs;
23 	unsigned int rm;
24 	u64 rm64;
25 	u64 zm64;
26 	int s;
27 
28 	COMPXSP;
29 	COMPYSP;
30 	COMPZSP;
31 
32 	EXPLODEXSP;
33 	EXPLODEYSP;
34 	EXPLODEZSP;
35 
36 	FLUSHXSP;
37 	FLUSHYSP;
38 	FLUSHZSP;
39 
40 	ieee754_clearcx();
41 
42 	/*
43 	 * Handle the cases when at least one of x, y or z is a NaN.
44 	 * Order of precedence is sNaN, qNaN and z, x, y.
45 	 */
46 	if (zc == IEEE754_CLASS_SNAN)
47 		return ieee754sp_nanxcpt(z);
48 	if (xc == IEEE754_CLASS_SNAN)
49 		return ieee754sp_nanxcpt(x);
50 	if (yc == IEEE754_CLASS_SNAN)
51 		return ieee754sp_nanxcpt(y);
52 	if (zc == IEEE754_CLASS_QNAN)
53 		return z;
54 	if (xc == IEEE754_CLASS_QNAN)
55 		return x;
56 	if (yc == IEEE754_CLASS_QNAN)
57 		return y;
58 
59 	if (zc == IEEE754_CLASS_DNORM)
60 		SPDNORMZ;
61 	/* ZERO z cases are handled separately below */
62 
63 	switch (CLPAIR(xc, yc)) {
64 
65 
66 	/*
67 	 * Infinity handling
68 	 */
69 	case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_ZERO):
70 	case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_INF):
71 		ieee754_setcx(IEEE754_INVALID_OPERATION);
72 		return ieee754sp_indef();
73 
74 	case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_INF):
75 	case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_INF):
76 	case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_NORM):
77 	case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_DNORM):
78 	case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_INF):
79 		if ((zc == IEEE754_CLASS_INF) &&
80 		    ((!(flags & MADDF_NEGATE_PRODUCT) && (zs != (xs ^ ys))) ||
81 		     ((flags & MADDF_NEGATE_PRODUCT) && (zs == (xs ^ ys))))) {
82 			/*
83 			 * Cases of addition of infinities with opposite signs
84 			 * or subtraction of infinities with same signs.
85 			 */
86 			ieee754_setcx(IEEE754_INVALID_OPERATION);
87 			return ieee754sp_indef();
88 		}
89 		/*
90 		 * z is here either not an infinity, or an infinity having the
91 		 * same sign as product (x*y) (in case of MADDF.D instruction)
92 		 * or product -(x*y) (in MSUBF.D case). The result must be an
93 		 * infinity, and its sign is determined only by the value of
94 		 * (flags & MADDF_NEGATE_PRODUCT) and the signs of x and y.
95 		 */
96 		if (flags & MADDF_NEGATE_PRODUCT)
97 			return ieee754sp_inf(1 ^ (xs ^ ys));
98 		else
99 			return ieee754sp_inf(xs ^ ys);
100 
101 	case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_ZERO):
102 	case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_NORM):
103 	case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_DNORM):
104 	case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_ZERO):
105 	case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_ZERO):
106 		if (zc == IEEE754_CLASS_INF)
107 			return ieee754sp_inf(zs);
108 		if (zc == IEEE754_CLASS_ZERO) {
109 			/* Handle cases +0 + (-0) and similar ones. */
110 			if ((!(flags & MADDF_NEGATE_PRODUCT)
111 					&& (zs == (xs ^ ys))) ||
112 			    ((flags & MADDF_NEGATE_PRODUCT)
113 					&& (zs != (xs ^ ys))))
114 				/*
115 				 * Cases of addition of zeros of equal signs
116 				 * or subtraction of zeroes of opposite signs.
117 				 * The sign of the resulting zero is in any
118 				 * such case determined only by the sign of z.
119 				 */
120 				return z;
121 
122 			return ieee754sp_zero(ieee754_csr.rm == FPU_CSR_RD);
123 		}
124 		/* x*y is here 0, and z is not 0, so just return z */
125 		return z;
126 
127 	case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_DNORM):
128 		SPDNORMX;
129 		/* fall through */
130 
131 	case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_DNORM):
132 		if (zc == IEEE754_CLASS_INF)
133 			return ieee754sp_inf(zs);
134 		SPDNORMY;
135 		break;
136 
137 	case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_NORM):
138 		if (zc == IEEE754_CLASS_INF)
139 			return ieee754sp_inf(zs);
140 		SPDNORMX;
141 		break;
142 
143 	case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_NORM):
144 		if (zc == IEEE754_CLASS_INF)
145 			return ieee754sp_inf(zs);
146 		/* continue to real computations */
147 	}
148 
149 	/* Finally get to do some computation */
150 
151 	/*
152 	 * Do the multiplication bit first
153 	 *
154 	 * rm = xm * ym, re = xe + ye basically
155 	 *
156 	 * At this point xm and ym should have been normalized.
157 	 */
158 
159 	/* rm = xm * ym, re = xe+ye basically */
160 	assert(xm & SP_HIDDEN_BIT);
161 	assert(ym & SP_HIDDEN_BIT);
162 
163 	re = xe + ye;
164 	rs = xs ^ ys;
165 	if (flags & MADDF_NEGATE_PRODUCT)
166 		rs ^= 1;
167 
168 	/* Multiple 24 bit xm and ym to give 48 bit results */
169 	rm64 = (uint64_t)xm * ym;
170 
171 	/* Shunt to top of word */
172 	rm64 = rm64 << 16;
173 
174 	/* Put explicit bit at bit 62 if necessary */
175 	if ((int64_t) rm64 < 0) {
176 		rm64 = rm64 >> 1;
177 		re++;
178 	}
179 
180 	assert(rm64 & (1 << 62));
181 
182 	if (zc == IEEE754_CLASS_ZERO) {
183 		/*
184 		 * Move explicit bit from bit 62 to bit 26 since the
185 		 * ieee754sp_format code expects the mantissa to be
186 		 * 27 bits wide (24 + 3 rounding bits).
187 		 */
188 		rm = XSPSRS64(rm64, (62 - 26));
189 		return ieee754sp_format(rs, re, rm);
190 	}
191 
192 	/* Move explicit bit from bit 23 to bit 62 */
193 	zm64 = (uint64_t)zm << (62 - 23);
194 	assert(zm64 & (1 << 62));
195 
196 	/* Make the exponents the same */
197 	if (ze > re) {
198 		/*
199 		 * Have to shift r fraction right to align.
200 		 */
201 		s = ze - re;
202 		rm64 = XSPSRS64(rm64, s);
203 		re += s;
204 	} else if (re > ze) {
205 		/*
206 		 * Have to shift z fraction right to align.
207 		 */
208 		s = re - ze;
209 		zm64 = XSPSRS64(zm64, s);
210 		ze += s;
211 	}
212 	assert(ze == re);
213 	assert(ze <= SP_EMAX);
214 
215 	/* Do the addition */
216 	if (zs == rs) {
217 		/*
218 		 * Generate 64 bit result by adding two 63 bit numbers
219 		 * leaving result in zm64, zs and ze.
220 		 */
221 		zm64 = zm64 + rm64;
222 		if ((int64_t)zm64 < 0) {	/* carry out */
223 			zm64 = XSPSRS1(zm64);
224 			ze++;
225 		}
226 	} else {
227 		if (zm64 >= rm64) {
228 			zm64 = zm64 - rm64;
229 		} else {
230 			zm64 = rm64 - zm64;
231 			zs = rs;
232 		}
233 		if (zm64 == 0)
234 			return ieee754sp_zero(ieee754_csr.rm == FPU_CSR_RD);
235 
236 		/*
237 		 * Put explicit bit at bit 62 if necessary.
238 		 */
239 		while ((zm64 >> 62) == 0) {
240 			zm64 <<= 1;
241 			ze--;
242 		}
243 	}
244 
245 	/*
246 	 * Move explicit bit from bit 62 to bit 26 since the
247 	 * ieee754sp_format code expects the mantissa to be
248 	 * 27 bits wide (24 + 3 rounding bits).
249 	 */
250 	zm = XSPSRS64(zm64, (62 - 26));
251 
252 	return ieee754sp_format(zs, ze, zm);
253 }
254 
255 union ieee754sp ieee754sp_maddf(union ieee754sp z, union ieee754sp x,
256 				union ieee754sp y)
257 {
258 	return _sp_maddf(z, x, y, 0);
259 }
260 
261 union ieee754sp ieee754sp_msubf(union ieee754sp z, union ieee754sp x,
262 				union ieee754sp y)
263 {
264 	return _sp_maddf(z, x, y, MADDF_NEGATE_PRODUCT);
265 }
266