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])
6 * MIPS floating point support
7 * Copyright (C) 2015 Imagination Technologies, Ltd.
8 * Author: Markos Chandras <markos.chandras@imgtec.com>
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.
15 #include "ieee754sp.h"
18 static union ieee754sp _sp_maddf(union ieee754sp z, union ieee754sp x,
19 union ieee754sp y, enum maddf_flags flags)
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.
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)
54 if (xc == IEEE754_CLASS_QNAN)
56 if (yc == IEEE754_CLASS_QNAN)
59 if (zc == IEEE754_CLASS_DNORM)
61 /* ZERO z cases are handled separately below */
63 switch (CLPAIR(xc, yc)) {
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();
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))))) {
83 * Cases of addition of infinities with opposite signs
84 * or subtraction of infinities with same signs.
86 ieee754_setcx(IEEE754_INVALID_OPERATION);
87 return ieee754sp_indef();
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.
96 if (flags & MADDF_NEGATE_PRODUCT)
97 return ieee754sp_inf(1 ^ (xs ^ ys));
99 return ieee754sp_inf(xs ^ ys);
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))))
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.
122 return ieee754sp_zero(ieee754_csr.rm == FPU_CSR_RD);
124 /* x*y is here 0, and z is not 0, so just return z */
127 case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_DNORM):
130 case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_DNORM):
131 if (zc == IEEE754_CLASS_INF)
132 return ieee754sp_inf(zs);
136 case CLPAIR(IEEE754_CLASS_DNORM, IEEE754_CLASS_NORM):
137 if (zc == IEEE754_CLASS_INF)
138 return ieee754sp_inf(zs);
142 case CLPAIR(IEEE754_CLASS_NORM, IEEE754_CLASS_NORM):
143 if (zc == IEEE754_CLASS_INF)
144 return ieee754sp_inf(zs);
145 /* fall through to real computations */
148 /* Finally get to do some computation */
151 * Do the multiplication bit first
153 * rm = xm * ym, re = xe + ye basically
155 * At this point xm and ym should have been normalized.
158 /* rm = xm * ym, re = xe+ye basically */
159 assert(xm & SP_HIDDEN_BIT);
160 assert(ym & SP_HIDDEN_BIT);
164 if (flags & MADDF_NEGATE_PRODUCT)
167 /* Multiple 24 bit xm and ym to give 48 bit results */
168 rm64 = (uint64_t)xm * ym;
170 /* Shunt to top of word */
173 /* Put explicit bit at bit 62 if necessary */
174 if ((int64_t) rm64 < 0) {
179 assert(rm64 & (1 << 62));
181 if (zc == IEEE754_CLASS_ZERO) {
183 * Move explicit bit from bit 62 to bit 26 since the
184 * ieee754sp_format code expects the mantissa to be
185 * 27 bits wide (24 + 3 rounding bits).
187 rm = XSPSRS64(rm64, (62 - 26));
188 return ieee754sp_format(rs, re, rm);
191 /* Move explicit bit from bit 23 to bit 62 */
192 zm64 = (uint64_t)zm << (62 - 23);
193 assert(zm64 & (1 << 62));
195 /* Make the exponents the same */
198 * Have to shift r fraction right to align.
201 rm64 = XSPSRS64(rm64, s);
203 } else if (re > ze) {
205 * Have to shift z fraction right to align.
208 zm64 = XSPSRS64(zm64, s);
212 assert(ze <= SP_EMAX);
214 /* Do the addition */
217 * Generate 64 bit result by adding two 63 bit numbers
218 * leaving result in zm64, zs and ze.
221 if ((int64_t)zm64 < 0) { /* carry out */
222 zm64 = XSPSRS1(zm64);
233 return ieee754sp_zero(ieee754_csr.rm == FPU_CSR_RD);
236 * Put explicit bit at bit 62 if necessary.
238 while ((zm64 >> 62) == 0) {
245 * Move explicit bit from bit 62 to bit 26 since the
246 * ieee754sp_format code expects the mantissa to be
247 * 27 bits wide (24 + 3 rounding bits).
249 zm = XSPSRS64(zm64, (62 - 26));
251 return ieee754sp_format(zs, ze, zm);
254 union ieee754sp ieee754sp_maddf(union ieee754sp z, union ieee754sp x,
257 return _sp_maddf(z, x, y, 0);
260 union ieee754sp ieee754sp_msubf(union ieee754sp z, union ieee754sp x,
263 return _sp_maddf(z, x, y, MADDF_NEGATE_PRODUCT);