xref: /openbmc/linux/arch/mips/math-emu/sp_maddf.c (revision 8795a739)
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 	/*
40 	 * Handle the cases when at least one of x, y or z is a NaN.
41 	 * Order of precedence is sNaN, qNaN and z, x, y.
42 	 */
43 	if (zc == IEEE754_CLASS_SNAN)
44 		return ieee754sp_nanxcpt(z);
45 	if (xc == IEEE754_CLASS_SNAN)
46 		return ieee754sp_nanxcpt(x);
47 	if (yc == IEEE754_CLASS_SNAN)
48 		return ieee754sp_nanxcpt(y);
49 	if (zc == IEEE754_CLASS_QNAN)
50 		return z;
51 	if (xc == IEEE754_CLASS_QNAN)
52 		return x;
53 	if (yc == IEEE754_CLASS_QNAN)
54 		return y;
55 
56 	if (zc == IEEE754_CLASS_DNORM)
57 		SPDNORMZ;
58 	/* ZERO z cases are handled separately below */
59 
60 	switch (CLPAIR(xc, yc)) {
61 
62 
63 	/*
64 	 * Infinity handling
65 	 */
66 	case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_ZERO):
67 	case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_INF):
68 		ieee754_setcx(IEEE754_INVALID_OPERATION);
69 		return ieee754sp_indef();
70 
71 	case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_INF):
72 	case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_INF):
73 	case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_NORM):
74 	case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_DNORM):
75 	case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_INF):
76 		if ((zc == IEEE754_CLASS_INF) &&
77 		    ((!(flags & MADDF_NEGATE_PRODUCT) && (zs != (xs ^ ys))) ||
78 		     ((flags & MADDF_NEGATE_PRODUCT) && (zs == (xs ^ ys))))) {
79 			/*
80 			 * Cases of addition of infinities with opposite signs
81 			 * or subtraction of infinities with same signs.
82 			 */
83 			ieee754_setcx(IEEE754_INVALID_OPERATION);
84 			return ieee754sp_indef();
85 		}
86 		/*
87 		 * z is here either not an infinity, or an infinity having the
88 		 * same sign as product (x*y) (in case of MADDF.D instruction)
89 		 * or product -(x*y) (in MSUBF.D case). The result must be an
90 		 * infinity, and its sign is determined only by the value of
91 		 * (flags & MADDF_NEGATE_PRODUCT) and the signs of x and y.
92 		 */
93 		if (flags & MADDF_NEGATE_PRODUCT)
94 			return ieee754sp_inf(1 ^ (xs ^ ys));
95 		else
96 			return ieee754sp_inf(xs ^ ys);
97 
98 	case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_ZERO):
99 	case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_NORM):
100 	case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_DNORM):
101 	case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_ZERO):
102 	case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_ZERO):
103 		if (zc == IEEE754_CLASS_INF)
104 			return ieee754sp_inf(zs);
105 		if (zc == IEEE754_CLASS_ZERO) {
106 			/* Handle cases +0 + (-0) and similar ones. */
107 			if ((!(flags & MADDF_NEGATE_PRODUCT)
108 					&& (zs == (xs ^ ys))) ||
109 			    ((flags & MADDF_NEGATE_PRODUCT)
110 					&& (zs != (xs ^ ys))))
111 				/*
112 				 * Cases of addition of zeros of equal signs
113 				 * or subtraction of zeroes of opposite signs.
114 				 * The sign of the resulting zero is in any
115 				 * such case determined only by the sign of z.
116 				 */
117 				return z;
118 
119 			return ieee754sp_zero(ieee754_csr.rm == FPU_CSR_RD);
120 		}
121 		/* x*y is here 0, and z is not 0, so just return z */
122 		return z;
123 
124 	case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_DNORM):
125 		SPDNORMX;
126 		/* fall through */
127 
128 	case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_DNORM):
129 		if (zc == IEEE754_CLASS_INF)
130 			return ieee754sp_inf(zs);
131 		SPDNORMY;
132 		break;
133 
134 	case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_NORM):
135 		if (zc == IEEE754_CLASS_INF)
136 			return ieee754sp_inf(zs);
137 		SPDNORMX;
138 		break;
139 
140 	case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_NORM):
141 		if (zc == IEEE754_CLASS_INF)
142 			return ieee754sp_inf(zs);
143 		/* continue to real computations */
144 	}
145 
146 	/* Finally get to do some computation */
147 
148 	/*
149 	 * Do the multiplication bit first
150 	 *
151 	 * rm = xm * ym, re = xe + ye basically
152 	 *
153 	 * At this point xm and ym should have been normalized.
154 	 */
155 
156 	/* rm = xm * ym, re = xe+ye basically */
157 	assert(xm & SP_HIDDEN_BIT);
158 	assert(ym & SP_HIDDEN_BIT);
159 
160 	re = xe + ye;
161 	rs = xs ^ ys;
162 	if (flags & MADDF_NEGATE_PRODUCT)
163 		rs ^= 1;
164 
165 	/* Multiple 24 bit xm and ym to give 48 bit results */
166 	rm64 = (uint64_t)xm * ym;
167 
168 	/* Shunt to top of word */
169 	rm64 = rm64 << 16;
170 
171 	/* Put explicit bit at bit 62 if necessary */
172 	if ((int64_t) rm64 < 0) {
173 		rm64 = rm64 >> 1;
174 		re++;
175 	}
176 
177 	assert(rm64 & (1 << 62));
178 
179 	if (zc == IEEE754_CLASS_ZERO) {
180 		/*
181 		 * Move explicit bit from bit 62 to bit 26 since the
182 		 * ieee754sp_format code expects the mantissa to be
183 		 * 27 bits wide (24 + 3 rounding bits).
184 		 */
185 		rm = XSPSRS64(rm64, (62 - 26));
186 		return ieee754sp_format(rs, re, rm);
187 	}
188 
189 	/* Move explicit bit from bit 23 to bit 62 */
190 	zm64 = (uint64_t)zm << (62 - 23);
191 	assert(zm64 & (1 << 62));
192 
193 	/* Make the exponents the same */
194 	if (ze > re) {
195 		/*
196 		 * Have to shift r fraction right to align.
197 		 */
198 		s = ze - re;
199 		rm64 = XSPSRS64(rm64, s);
200 		re += s;
201 	} else if (re > ze) {
202 		/*
203 		 * Have to shift z fraction right to align.
204 		 */
205 		s = re - ze;
206 		zm64 = XSPSRS64(zm64, s);
207 		ze += s;
208 	}
209 	assert(ze == re);
210 	assert(ze <= SP_EMAX);
211 
212 	/* Do the addition */
213 	if (zs == rs) {
214 		/*
215 		 * Generate 64 bit result by adding two 63 bit numbers
216 		 * leaving result in zm64, zs and ze.
217 		 */
218 		zm64 = zm64 + rm64;
219 		if ((int64_t)zm64 < 0) {	/* carry out */
220 			zm64 = XSPSRS1(zm64);
221 			ze++;
222 		}
223 	} else {
224 		if (zm64 >= rm64) {
225 			zm64 = zm64 - rm64;
226 		} else {
227 			zm64 = rm64 - zm64;
228 			zs = rs;
229 		}
230 		if (zm64 == 0)
231 			return ieee754sp_zero(ieee754_csr.rm == FPU_CSR_RD);
232 
233 		/*
234 		 * Put explicit bit at bit 62 if necessary.
235 		 */
236 		while ((zm64 >> 62) == 0) {
237 			zm64 <<= 1;
238 			ze--;
239 		}
240 	}
241 
242 	/*
243 	 * Move explicit bit from bit 62 to bit 26 since the
244 	 * ieee754sp_format code expects the mantissa to be
245 	 * 27 bits wide (24 + 3 rounding bits).
246 	 */
247 	zm = XSPSRS64(zm64, (62 - 26));
248 
249 	return ieee754sp_format(zs, ze, zm);
250 }
251 
252 union ieee754sp ieee754sp_maddf(union ieee754sp z, union ieee754sp x,
253 				union ieee754sp y)
254 {
255 	return _sp_maddf(z, x, y, 0);
256 }
257 
258 union ieee754sp ieee754sp_msubf(union ieee754sp z, union ieee754sp x,
259 				union ieee754sp y)
260 {
261 	return _sp_maddf(z, x, y, MADDF_NEGATE_PRODUCT);
262 }
263