arch/mips/math-emu/sp_maddf.c

   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 }