xref: /openbmc/linux/arch/mips/math-emu/dp_maddf.c (revision b78412b8)
1 /*
2  * IEEE754 floating point arithmetic
3  * double 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 "ieee754dp.h"
16 
17 
18 /* 128 bits shift right logical with rounding. */
19 void srl128(u64 *hptr, u64 *lptr, int count)
20 {
21 	u64 low;
22 
23 	if (count >= 128) {
24 		*lptr = *hptr != 0 || *lptr != 0;
25 		*hptr = 0;
26 	} else if (count >= 64) {
27 		if (count == 64) {
28 			*lptr = *hptr | (*lptr != 0);
29 		} else {
30 			low = *lptr;
31 			*lptr = *hptr >> (count - 64);
32 			*lptr |= (*hptr << (128 - count)) != 0 || low != 0;
33 		}
34 		*hptr = 0;
35 	} else {
36 		low = *lptr;
37 		*lptr = low >> count | *hptr << (64 - count);
38 		*lptr |= (low << (64 - count)) != 0;
39 		*hptr = *hptr >> count;
40 	}
41 }
42 
43 static union ieee754dp _dp_maddf(union ieee754dp z, union ieee754dp x,
44 				 union ieee754dp y, enum maddf_flags flags)
45 {
46 	int re;
47 	int rs;
48 	unsigned lxm;
49 	unsigned hxm;
50 	unsigned lym;
51 	unsigned hym;
52 	u64 lrm;
53 	u64 hrm;
54 	u64 lzm;
55 	u64 hzm;
56 	u64 t;
57 	u64 at;
58 	int s;
59 
60 	COMPXDP;
61 	COMPYDP;
62 	COMPZDP;
63 
64 	EXPLODEXDP;
65 	EXPLODEYDP;
66 	EXPLODEZDP;
67 
68 	FLUSHXDP;
69 	FLUSHYDP;
70 	FLUSHZDP;
71 
72 	ieee754_clearcx();
73 
74 	/*
75 	 * Handle the cases when at least one of x, y or z is a NaN.
76 	 * Order of precedence is sNaN, qNaN and z, x, y.
77 	 */
78 	if (zc == IEEE754_CLASS_SNAN)
79 		return ieee754dp_nanxcpt(z);
80 	if (xc == IEEE754_CLASS_SNAN)
81 		return ieee754dp_nanxcpt(x);
82 	if (yc == IEEE754_CLASS_SNAN)
83 		return ieee754dp_nanxcpt(y);
84 	if (zc == IEEE754_CLASS_QNAN)
85 		return z;
86 	if (xc == IEEE754_CLASS_QNAN)
87 		return x;
88 	if (yc == IEEE754_CLASS_QNAN)
89 		return y;
90 
91 	if (zc == IEEE754_CLASS_DNORM)
92 		DPDNORMZ;
93 	/* ZERO z cases are handled separately below */
94 
95 	switch (CLPAIR(xc, yc)) {
96 
97 	/*
98 	 * Infinity handling
99 	 */
100 	case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_ZERO):
101 	case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_INF):
102 		ieee754_setcx(IEEE754_INVALID_OPERATION);
103 		return ieee754dp_indef();
104 
105 	case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_INF):
106 	case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_INF):
107 	case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_NORM):
108 	case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_DNORM):
109 	case CLPAIR(IEEE754_CLASS_INF, IEEE754_CLASS_INF):
110 		if ((zc == IEEE754_CLASS_INF) &&
111 		    ((!(flags & MADDF_NEGATE_PRODUCT) && (zs != (xs ^ ys))) ||
112 		     ((flags & MADDF_NEGATE_PRODUCT) && (zs == (xs ^ ys))))) {
113 			/*
114 			 * Cases of addition of infinities with opposite signs
115 			 * or subtraction of infinities with same signs.
116 			 */
117 			ieee754_setcx(IEEE754_INVALID_OPERATION);
118 			return ieee754dp_indef();
119 		}
120 		/*
121 		 * z is here either not an infinity, or an infinity having the
122 		 * same sign as product (x*y) (in case of MADDF.D instruction)
123 		 * or product -(x*y) (in MSUBF.D case). The result must be an
124 		 * infinity, and its sign is determined only by the value of
125 		 * (flags & MADDF_NEGATE_PRODUCT) and the signs of x and y.
126 		 */
127 		if (flags & MADDF_NEGATE_PRODUCT)
128 			return ieee754dp_inf(1 ^ (xs ^ ys));
129 		else
130 			return ieee754dp_inf(xs ^ ys);
131 
132 	case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_ZERO):
133 	case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_NORM):
134 	case CLPAIR(IEEE754_CLASS_ZERO, IEEE754_CLASS_DNORM):
135 	case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_ZERO):
136 	case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_ZERO):
137 		if (zc == IEEE754_CLASS_INF)
138 			return ieee754dp_inf(zs);
139 		if (zc == IEEE754_CLASS_ZERO) {
140 			/* Handle cases +0 + (-0) and similar ones. */
141 			if ((!(flags & MADDF_NEGATE_PRODUCT)
142 					&& (zs == (xs ^ ys))) ||
143 			    ((flags & MADDF_NEGATE_PRODUCT)
144 					&& (zs != (xs ^ ys))))
145 				/*
146 				 * Cases of addition of zeros of equal signs
147 				 * or subtraction of zeroes of opposite signs.
148 				 * The sign of the resulting zero is in any
149 				 * such case determined only by the sign of z.
150 				 */
151 				return z;
152 
153 			return ieee754dp_zero(ieee754_csr.rm == FPU_CSR_RD);
154 		}
155 		/* x*y is here 0, and z is not 0, so just return z */
156 		return z;
157 
158 	case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_DNORM):
159 		DPDNORMX;
160 
161 	case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_DNORM):
162 		if (zc == IEEE754_CLASS_INF)
163 			return ieee754dp_inf(zs);
164 		DPDNORMY;
165 		break;
166 
167 	case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_NORM):
168 		if (zc == IEEE754_CLASS_INF)
169 			return ieee754dp_inf(zs);
170 		DPDNORMX;
171 		break;
172 
173 	case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_NORM):
174 		if (zc == IEEE754_CLASS_INF)
175 			return ieee754dp_inf(zs);
176 		/* fall through to real computations */
177 	}
178 
179 	/* Finally get to do some computation */
180 
181 	/*
182 	 * Do the multiplication bit first
183 	 *
184 	 * rm = xm * ym, re = xe + ye basically
185 	 *
186 	 * At this point xm and ym should have been normalized.
187 	 */
188 	assert(xm & DP_HIDDEN_BIT);
189 	assert(ym & DP_HIDDEN_BIT);
190 
191 	re = xe + ye;
192 	rs = xs ^ ys;
193 	if (flags & MADDF_NEGATE_PRODUCT)
194 		rs ^= 1;
195 
196 	/* shunt to top of word */
197 	xm <<= 64 - (DP_FBITS + 1);
198 	ym <<= 64 - (DP_FBITS + 1);
199 
200 	/*
201 	 * Multiply 64 bits xm and ym to give 128 bits result in hrm:lrm.
202 	 */
203 
204 	/* 32 * 32 => 64 */
205 #define DPXMULT(x, y)	((u64)(x) * (u64)y)
206 
207 	lxm = xm;
208 	hxm = xm >> 32;
209 	lym = ym;
210 	hym = ym >> 32;
211 
212 	lrm = DPXMULT(lxm, lym);
213 	hrm = DPXMULT(hxm, hym);
214 
215 	t = DPXMULT(lxm, hym);
216 
217 	at = lrm + (t << 32);
218 	hrm += at < lrm;
219 	lrm = at;
220 
221 	hrm = hrm + (t >> 32);
222 
223 	t = DPXMULT(hxm, lym);
224 
225 	at = lrm + (t << 32);
226 	hrm += at < lrm;
227 	lrm = at;
228 
229 	hrm = hrm + (t >> 32);
230 
231 	/* Put explicit bit at bit 126 if necessary */
232 	if ((int64_t)hrm < 0) {
233 		lrm = (hrm << 63) | (lrm >> 1);
234 		hrm = hrm >> 1;
235 		re++;
236 	}
237 
238 	assert(hrm & (1 << 62));
239 
240 	if (zc == IEEE754_CLASS_ZERO) {
241 		/*
242 		 * Move explicit bit from bit 126 to bit 55 since the
243 		 * ieee754dp_format code expects the mantissa to be
244 		 * 56 bits wide (53 + 3 rounding bits).
245 		 */
246 		srl128(&hrm, &lrm, (126 - 55));
247 		return ieee754dp_format(rs, re, lrm);
248 	}
249 
250 	/* Move explicit bit from bit 52 to bit 126 */
251 	lzm = 0;
252 	hzm = zm << 10;
253 	assert(hzm & (1 << 62));
254 
255 	/* Make the exponents the same */
256 	if (ze > re) {
257 		/*
258 		 * Have to shift y fraction right to align.
259 		 */
260 		s = ze - re;
261 		srl128(&hrm, &lrm, s);
262 		re += s;
263 	} else if (re > ze) {
264 		/*
265 		 * Have to shift x fraction right to align.
266 		 */
267 		s = re - ze;
268 		srl128(&hzm, &lzm, s);
269 		ze += s;
270 	}
271 	assert(ze == re);
272 	assert(ze <= DP_EMAX);
273 
274 	/* Do the addition */
275 	if (zs == rs) {
276 		/*
277 		 * Generate 128 bit result by adding two 127 bit numbers
278 		 * leaving result in hzm:lzm, zs and ze.
279 		 */
280 		hzm = hzm + hrm + (lzm > (lzm + lrm));
281 		lzm = lzm + lrm;
282 		if ((int64_t)hzm < 0) {        /* carry out */
283 			srl128(&hzm, &lzm, 1);
284 			ze++;
285 		}
286 	} else {
287 		if (hzm > hrm || (hzm == hrm && lzm >= lrm)) {
288 			hzm = hzm - hrm - (lzm < lrm);
289 			lzm = lzm - lrm;
290 		} else {
291 			hzm = hrm - hzm - (lrm < lzm);
292 			lzm = lrm - lzm;
293 			zs = rs;
294 		}
295 		if (lzm == 0 && hzm == 0)
296 			return ieee754dp_zero(ieee754_csr.rm == FPU_CSR_RD);
297 
298 		/*
299 		 * Put explicit bit at bit 126 if necessary.
300 		 */
301 		if (hzm == 0) {
302 			/* left shift by 63 or 64 bits */
303 			if ((int64_t)lzm < 0) {
304 				/* MSB of lzm is the explicit bit */
305 				hzm = lzm >> 1;
306 				lzm = lzm << 63;
307 				ze -= 63;
308 			} else {
309 				hzm = lzm;
310 				lzm = 0;
311 				ze -= 64;
312 			}
313 		}
314 
315 		t = 0;
316 		while ((hzm >> (62 - t)) == 0)
317 			t++;
318 
319 		assert(t <= 62);
320 		if (t) {
321 			hzm = hzm << t | lzm >> (64 - t);
322 			lzm = lzm << t;
323 			ze -= t;
324 		}
325 	}
326 
327 	/*
328 	 * Move explicit bit from bit 126 to bit 55 since the
329 	 * ieee754dp_format code expects the mantissa to be
330 	 * 56 bits wide (53 + 3 rounding bits).
331 	 */
332 	srl128(&hzm, &lzm, (126 - 55));
333 
334 	return ieee754dp_format(zs, ze, lzm);
335 }
336 
337 union ieee754dp ieee754dp_maddf(union ieee754dp z, union ieee754dp x,
338 				union ieee754dp y)
339 {
340 	return _dp_maddf(z, x, y, 0);
341 }
342 
343 union ieee754dp ieee754dp_msubf(union ieee754dp z, union ieee754dp x,
344 				union ieee754dp y)
345 {
346 	return _dp_maddf(z, x, y, MADDF_NEGATE_PRODUCT);
347 }
348