| 1 |
3 |
xianfeng |
#define _FP_DECL(wc, X) \
|
| 2 |
|
|
_FP_I_TYPE X##_c, X##_s, X##_e; \
|
| 3 |
|
|
_FP_FRAC_DECL_##wc(X)
|
| 4 |
|
|
|
| 5 |
|
|
/*
|
| 6 |
|
|
* Finish truely unpacking a native fp value by classifying the kind
|
| 7 |
|
|
* of fp value and normalizing both the exponent and the fraction.
|
| 8 |
|
|
*/
|
| 9 |
|
|
|
| 10 |
|
|
#define _FP_UNPACK_CANONICAL(fs, wc, X) \
|
| 11 |
|
|
do { \
|
| 12 |
|
|
switch (X##_e) \
|
| 13 |
|
|
{ \
|
| 14 |
|
|
default: \
|
| 15 |
|
|
_FP_FRAC_HIGH_##wc(X) |= _FP_IMPLBIT_##fs; \
|
| 16 |
|
|
_FP_FRAC_SLL_##wc(X, _FP_WORKBITS); \
|
| 17 |
|
|
X##_e -= _FP_EXPBIAS_##fs; \
|
| 18 |
|
|
X##_c = FP_CLS_NORMAL; \
|
| 19 |
|
|
break; \
|
| 20 |
|
|
\
|
| 21 |
|
|
case 0: \
|
| 22 |
|
|
if (_FP_FRAC_ZEROP_##wc(X)) \
|
| 23 |
|
|
X##_c = FP_CLS_ZERO; \
|
| 24 |
|
|
else \
|
| 25 |
|
|
{ \
|
| 26 |
|
|
/* a denormalized number */ \
|
| 27 |
|
|
_FP_I_TYPE _shift; \
|
| 28 |
|
|
_FP_FRAC_CLZ_##wc(_shift, X); \
|
| 29 |
|
|
_shift -= _FP_FRACXBITS_##fs; \
|
| 30 |
|
|
_FP_FRAC_SLL_##wc(X, (_shift+_FP_WORKBITS)); \
|
| 31 |
|
|
X##_e -= _FP_EXPBIAS_##fs - 1 + _shift; \
|
| 32 |
|
|
X##_c = FP_CLS_NORMAL; \
|
| 33 |
|
|
} \
|
| 34 |
|
|
break; \
|
| 35 |
|
|
\
|
| 36 |
|
|
case _FP_EXPMAX_##fs: \
|
| 37 |
|
|
if (_FP_FRAC_ZEROP_##wc(X)) \
|
| 38 |
|
|
X##_c = FP_CLS_INF; \
|
| 39 |
|
|
else \
|
| 40 |
|
|
/* we don't differentiate between signaling and quiet nans */ \
|
| 41 |
|
|
X##_c = FP_CLS_NAN; \
|
| 42 |
|
|
break; \
|
| 43 |
|
|
} \
|
| 44 |
|
|
} while (0)
|
| 45 |
|
|
|
| 46 |
|
|
|
| 47 |
|
|
/*
|
| 48 |
|
|
* Before packing the bits back into the native fp result, take care
|
| 49 |
|
|
* of such mundane things as rounding and overflow. Also, for some
|
| 50 |
|
|
* kinds of fp values, the original parts may not have been fully
|
| 51 |
|
|
* extracted -- but that is ok, we can regenerate them now.
|
| 52 |
|
|
*/
|
| 53 |
|
|
|
| 54 |
|
|
#define _FP_PACK_CANONICAL(fs, wc, X) \
|
| 55 |
|
|
({int __ret = 0; \
|
| 56 |
|
|
switch (X##_c) \
|
| 57 |
|
|
{ \
|
| 58 |
|
|
case FP_CLS_NORMAL: \
|
| 59 |
|
|
X##_e += _FP_EXPBIAS_##fs; \
|
| 60 |
|
|
if (X##_e > 0) \
|
| 61 |
|
|
{ \
|
| 62 |
|
|
__ret |= _FP_ROUND(wc, X); \
|
| 63 |
|
|
if (_FP_FRAC_OVERP_##wc(fs, X)) \
|
| 64 |
|
|
{ \
|
| 65 |
|
|
_FP_FRAC_SRL_##wc(X, (_FP_WORKBITS+1)); \
|
| 66 |
|
|
X##_e++; \
|
| 67 |
|
|
} \
|
| 68 |
|
|
else \
|
| 69 |
|
|
_FP_FRAC_SRL_##wc(X, _FP_WORKBITS); \
|
| 70 |
|
|
if (X##_e >= _FP_EXPMAX_##fs) \
|
| 71 |
|
|
{ \
|
| 72 |
|
|
/* overflow to infinity */ \
|
| 73 |
|
|
X##_e = _FP_EXPMAX_##fs; \
|
| 74 |
|
|
_FP_FRAC_SET_##wc(X, _FP_ZEROFRAC_##wc); \
|
| 75 |
|
|
__ret |= EFLAG_OVERFLOW; \
|
| 76 |
|
|
} \
|
| 77 |
|
|
} \
|
| 78 |
|
|
else \
|
| 79 |
|
|
{ \
|
| 80 |
|
|
/* we've got a denormalized number */ \
|
| 81 |
|
|
X##_e = -X##_e + 1; \
|
| 82 |
|
|
if (X##_e <= _FP_WFRACBITS_##fs) \
|
| 83 |
|
|
{ \
|
| 84 |
|
|
_FP_FRAC_SRS_##wc(X, X##_e, _FP_WFRACBITS_##fs); \
|
| 85 |
|
|
_FP_FRAC_SLL_##wc(X, 1); \
|
| 86 |
|
|
if (_FP_FRAC_OVERP_##wc(fs, X)) \
|
| 87 |
|
|
{ \
|
| 88 |
|
|
X##_e = 1; \
|
| 89 |
|
|
_FP_FRAC_SET_##wc(X, _FP_ZEROFRAC_##wc); \
|
| 90 |
|
|
} \
|
| 91 |
|
|
else \
|
| 92 |
|
|
{ \
|
| 93 |
|
|
X##_e = 0; \
|
| 94 |
|
|
_FP_FRAC_SRL_##wc(X, _FP_WORKBITS+1); \
|
| 95 |
|
|
__ret |= EFLAG_UNDERFLOW; \
|
| 96 |
|
|
} \
|
| 97 |
|
|
} \
|
| 98 |
|
|
else \
|
| 99 |
|
|
{ \
|
| 100 |
|
|
/* underflow to zero */ \
|
| 101 |
|
|
X##_e = 0; \
|
| 102 |
|
|
_FP_FRAC_SET_##wc(X, _FP_ZEROFRAC_##wc); \
|
| 103 |
|
|
__ret |= EFLAG_UNDERFLOW; \
|
| 104 |
|
|
} \
|
| 105 |
|
|
} \
|
| 106 |
|
|
break; \
|
| 107 |
|
|
\
|
| 108 |
|
|
case FP_CLS_ZERO: \
|
| 109 |
|
|
X##_e = 0; \
|
| 110 |
|
|
_FP_FRAC_SET_##wc(X, _FP_ZEROFRAC_##wc); \
|
| 111 |
|
|
break; \
|
| 112 |
|
|
\
|
| 113 |
|
|
case FP_CLS_INF: \
|
| 114 |
|
|
X##_e = _FP_EXPMAX_##fs; \
|
| 115 |
|
|
_FP_FRAC_SET_##wc(X, _FP_ZEROFRAC_##wc); \
|
| 116 |
|
|
break; \
|
| 117 |
|
|
\
|
| 118 |
|
|
case FP_CLS_NAN: \
|
| 119 |
|
|
X##_e = _FP_EXPMAX_##fs; \
|
| 120 |
|
|
if (!_FP_KEEPNANFRACP) \
|
| 121 |
|
|
{ \
|
| 122 |
|
|
_FP_FRAC_SET_##wc(X, _FP_NANFRAC_##fs); \
|
| 123 |
|
|
X##_s = 0; \
|
| 124 |
|
|
} \
|
| 125 |
|
|
else \
|
| 126 |
|
|
_FP_FRAC_HIGH_##wc(X) |= _FP_QNANBIT_##fs; \
|
| 127 |
|
|
break; \
|
| 128 |
|
|
} \
|
| 129 |
|
|
__ret; \
|
| 130 |
|
|
})
|
| 131 |
|
|
|
| 132 |
|
|
|
| 133 |
|
|
/*
|
| 134 |
|
|
* Main addition routine. The input values should be cooked.
|
| 135 |
|
|
*/
|
| 136 |
|
|
|
| 137 |
|
|
#define _FP_ADD(fs, wc, R, X, Y) \
|
| 138 |
|
|
do { \
|
| 139 |
|
|
switch (_FP_CLS_COMBINE(X##_c, Y##_c)) \
|
| 140 |
|
|
{ \
|
| 141 |
|
|
case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_NORMAL): \
|
| 142 |
|
|
{ \
|
| 143 |
|
|
/* shift the smaller number so that its exponent matches the larger */ \
|
| 144 |
|
|
_FP_I_TYPE diff = X##_e - Y##_e; \
|
| 145 |
|
|
\
|
| 146 |
|
|
if (diff < 0) \
|
| 147 |
|
|
{ \
|
| 148 |
|
|
diff = -diff; \
|
| 149 |
|
|
if (diff <= _FP_WFRACBITS_##fs) \
|
| 150 |
|
|
_FP_FRAC_SRS_##wc(X, diff, _FP_WFRACBITS_##fs); \
|
| 151 |
|
|
else if (!_FP_FRAC_ZEROP_##wc(X)) \
|
| 152 |
|
|
_FP_FRAC_SET_##wc(X, _FP_MINFRAC_##wc); \
|
| 153 |
|
|
else \
|
| 154 |
|
|
_FP_FRAC_SET_##wc(X, _FP_ZEROFRAC_##wc); \
|
| 155 |
|
|
R##_e = Y##_e; \
|
| 156 |
|
|
} \
|
| 157 |
|
|
else \
|
| 158 |
|
|
{ \
|
| 159 |
|
|
if (diff > 0) \
|
| 160 |
|
|
{ \
|
| 161 |
|
|
if (diff <= _FP_WFRACBITS_##fs) \
|
| 162 |
|
|
_FP_FRAC_SRS_##wc(Y, diff, _FP_WFRACBITS_##fs); \
|
| 163 |
|
|
else if (!_FP_FRAC_ZEROP_##wc(Y)) \
|
| 164 |
|
|
_FP_FRAC_SET_##wc(Y, _FP_MINFRAC_##wc); \
|
| 165 |
|
|
else \
|
| 166 |
|
|
_FP_FRAC_SET_##wc(Y, _FP_ZEROFRAC_##wc); \
|
| 167 |
|
|
} \
|
| 168 |
|
|
R##_e = X##_e; \
|
| 169 |
|
|
} \
|
| 170 |
|
|
\
|
| 171 |
|
|
R##_c = FP_CLS_NORMAL; \
|
| 172 |
|
|
\
|
| 173 |
|
|
if (X##_s == Y##_s) \
|
| 174 |
|
|
{ \
|
| 175 |
|
|
R##_s = X##_s; \
|
| 176 |
|
|
_FP_FRAC_ADD_##wc(R, X, Y); \
|
| 177 |
|
|
if (_FP_FRAC_OVERP_##wc(fs, R)) \
|
| 178 |
|
|
{ \
|
| 179 |
|
|
_FP_FRAC_SRS_##wc(R, 1, _FP_WFRACBITS_##fs); \
|
| 180 |
|
|
R##_e++; \
|
| 181 |
|
|
} \
|
| 182 |
|
|
} \
|
| 183 |
|
|
else \
|
| 184 |
|
|
{ \
|
| 185 |
|
|
R##_s = X##_s; \
|
| 186 |
|
|
_FP_FRAC_SUB_##wc(R, X, Y); \
|
| 187 |
|
|
if (_FP_FRAC_ZEROP_##wc(R)) \
|
| 188 |
|
|
{ \
|
| 189 |
|
|
/* return an exact zero */ \
|
| 190 |
|
|
if (FP_ROUNDMODE == FP_RND_MINF) \
|
| 191 |
|
|
R##_s |= Y##_s; \
|
| 192 |
|
|
else \
|
| 193 |
|
|
R##_s &= Y##_s; \
|
| 194 |
|
|
R##_c = FP_CLS_ZERO; \
|
| 195 |
|
|
} \
|
| 196 |
|
|
else \
|
| 197 |
|
|
{ \
|
| 198 |
|
|
if (_FP_FRAC_NEGP_##wc(R)) \
|
| 199 |
|
|
{ \
|
| 200 |
|
|
_FP_FRAC_SUB_##wc(R, Y, X); \
|
| 201 |
|
|
R##_s = Y##_s; \
|
| 202 |
|
|
} \
|
| 203 |
|
|
\
|
| 204 |
|
|
/* renormalize after subtraction */ \
|
| 205 |
|
|
_FP_FRAC_CLZ_##wc(diff, R); \
|
| 206 |
|
|
diff -= _FP_WFRACXBITS_##fs; \
|
| 207 |
|
|
if (diff) \
|
| 208 |
|
|
{ \
|
| 209 |
|
|
R##_e -= diff; \
|
| 210 |
|
|
_FP_FRAC_SLL_##wc(R, diff); \
|
| 211 |
|
|
} \
|
| 212 |
|
|
} \
|
| 213 |
|
|
} \
|
| 214 |
|
|
break; \
|
| 215 |
|
|
} \
|
| 216 |
|
|
\
|
| 217 |
|
|
case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_NAN): \
|
| 218 |
|
|
_FP_CHOOSENAN(fs, wc, R, X, Y); \
|
| 219 |
|
|
break; \
|
| 220 |
|
|
\
|
| 221 |
|
|
case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_ZERO): \
|
| 222 |
|
|
R##_e = X##_e; \
|
| 223 |
|
|
case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_NORMAL): \
|
| 224 |
|
|
case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_INF): \
|
| 225 |
|
|
case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_ZERO): \
|
| 226 |
|
|
_FP_FRAC_COPY_##wc(R, X); \
|
| 227 |
|
|
R##_s = X##_s; \
|
| 228 |
|
|
R##_c = X##_c; \
|
| 229 |
|
|
break; \
|
| 230 |
|
|
\
|
| 231 |
|
|
case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_NORMAL): \
|
| 232 |
|
|
R##_e = Y##_e; \
|
| 233 |
|
|
case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_NAN): \
|
| 234 |
|
|
case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_NAN): \
|
| 235 |
|
|
case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_NAN): \
|
| 236 |
|
|
_FP_FRAC_COPY_##wc(R, Y); \
|
| 237 |
|
|
R##_s = Y##_s; \
|
| 238 |
|
|
R##_c = Y##_c; \
|
| 239 |
|
|
break; \
|
| 240 |
|
|
\
|
| 241 |
|
|
case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_INF): \
|
| 242 |
|
|
if (X##_s != Y##_s) \
|
| 243 |
|
|
{ \
|
| 244 |
|
|
/* +INF + -INF => NAN */ \
|
| 245 |
|
|
_FP_FRAC_SET_##wc(R, _FP_NANFRAC_##fs); \
|
| 246 |
|
|
R##_s = X##_s ^ Y##_s; \
|
| 247 |
|
|
R##_c = FP_CLS_NAN; \
|
| 248 |
|
|
break; \
|
| 249 |
|
|
} \
|
| 250 |
|
|
/* FALLTHRU */ \
|
| 251 |
|
|
\
|
| 252 |
|
|
case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_NORMAL): \
|
| 253 |
|
|
case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_ZERO): \
|
| 254 |
|
|
R##_s = X##_s; \
|
| 255 |
|
|
R##_c = FP_CLS_INF; \
|
| 256 |
|
|
break; \
|
| 257 |
|
|
\
|
| 258 |
|
|
case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_INF): \
|
| 259 |
|
|
case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_INF): \
|
| 260 |
|
|
R##_s = Y##_s; \
|
| 261 |
|
|
R##_c = FP_CLS_INF; \
|
| 262 |
|
|
break; \
|
| 263 |
|
|
\
|
| 264 |
|
|
case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_ZERO): \
|
| 265 |
|
|
/* make sure the sign is correct */ \
|
| 266 |
|
|
if (FP_ROUNDMODE == FP_RND_MINF) \
|
| 267 |
|
|
R##_s = X##_s | Y##_s; \
|
| 268 |
|
|
else \
|
| 269 |
|
|
R##_s = X##_s & Y##_s; \
|
| 270 |
|
|
R##_c = FP_CLS_ZERO; \
|
| 271 |
|
|
break; \
|
| 272 |
|
|
\
|
| 273 |
|
|
default: \
|
| 274 |
|
|
abort(); \
|
| 275 |
|
|
} \
|
| 276 |
|
|
} while (0)
|
| 277 |
|
|
|
| 278 |
|
|
|
| 279 |
|
|
/*
|
| 280 |
|
|
* Main negation routine. FIXME -- when we care about setting exception
|
| 281 |
|
|
* bits reliably, this will not do. We should examine all of the fp classes.
|
| 282 |
|
|
*/
|
| 283 |
|
|
|
| 284 |
|
|
#define _FP_NEG(fs, wc, R, X) \
|
| 285 |
|
|
do { \
|
| 286 |
|
|
_FP_FRAC_COPY_##wc(R, X); \
|
| 287 |
|
|
R##_c = X##_c; \
|
| 288 |
|
|
R##_e = X##_e; \
|
| 289 |
|
|
R##_s = 1 ^ X##_s; \
|
| 290 |
|
|
} while (0)
|
| 291 |
|
|
|
| 292 |
|
|
|
| 293 |
|
|
/*
|
| 294 |
|
|
* Main multiplication routine. The input values should be cooked.
|
| 295 |
|
|
*/
|
| 296 |
|
|
|
| 297 |
|
|
#define _FP_MUL(fs, wc, R, X, Y) \
|
| 298 |
|
|
do { \
|
| 299 |
|
|
R##_s = X##_s ^ Y##_s; \
|
| 300 |
|
|
switch (_FP_CLS_COMBINE(X##_c, Y##_c)) \
|
| 301 |
|
|
{ \
|
| 302 |
|
|
case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_NORMAL): \
|
| 303 |
|
|
R##_c = FP_CLS_NORMAL; \
|
| 304 |
|
|
R##_e = X##_e + Y##_e + 1; \
|
| 305 |
|
|
\
|
| 306 |
|
|
_FP_MUL_MEAT_##fs(R,X,Y); \
|
| 307 |
|
|
\
|
| 308 |
|
|
if (_FP_FRAC_OVERP_##wc(fs, R)) \
|
| 309 |
|
|
_FP_FRAC_SRS_##wc(R, 1, _FP_WFRACBITS_##fs); \
|
| 310 |
|
|
else \
|
| 311 |
|
|
R##_e--; \
|
| 312 |
|
|
break; \
|
| 313 |
|
|
\
|
| 314 |
|
|
case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_NAN): \
|
| 315 |
|
|
_FP_CHOOSENAN(fs, wc, R, X, Y); \
|
| 316 |
|
|
break; \
|
| 317 |
|
|
\
|
| 318 |
|
|
case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_NORMAL): \
|
| 319 |
|
|
case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_INF): \
|
| 320 |
|
|
case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_ZERO): \
|
| 321 |
|
|
R##_s = X##_s; \
|
| 322 |
|
|
\
|
| 323 |
|
|
case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_INF): \
|
| 324 |
|
|
case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_NORMAL): \
|
| 325 |
|
|
case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_NORMAL): \
|
| 326 |
|
|
case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_ZERO): \
|
| 327 |
|
|
_FP_FRAC_COPY_##wc(R, X); \
|
| 328 |
|
|
R##_c = X##_c; \
|
| 329 |
|
|
break; \
|
| 330 |
|
|
\
|
| 331 |
|
|
case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_NAN): \
|
| 332 |
|
|
case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_NAN): \
|
| 333 |
|
|
case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_NAN): \
|
| 334 |
|
|
R##_s = Y##_s; \
|
| 335 |
|
|
\
|
| 336 |
|
|
case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_INF): \
|
| 337 |
|
|
case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_ZERO): \
|
| 338 |
|
|
_FP_FRAC_COPY_##wc(R, Y); \
|
| 339 |
|
|
R##_c = Y##_c; \
|
| 340 |
|
|
break; \
|
| 341 |
|
|
\
|
| 342 |
|
|
case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_ZERO): \
|
| 343 |
|
|
case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_INF): \
|
| 344 |
|
|
R##_c = FP_CLS_NAN; \
|
| 345 |
|
|
_FP_FRAC_SET_##wc(R, _FP_NANFRAC_##fs); \
|
| 346 |
|
|
break; \
|
| 347 |
|
|
\
|
| 348 |
|
|
default: \
|
| 349 |
|
|
abort(); \
|
| 350 |
|
|
} \
|
| 351 |
|
|
} while (0)
|
| 352 |
|
|
|
| 353 |
|
|
|
| 354 |
|
|
/*
|
| 355 |
|
|
* Main division routine. The input values should be cooked.
|
| 356 |
|
|
*/
|
| 357 |
|
|
|
| 358 |
|
|
#define _FP_DIV(fs, wc, R, X, Y) \
|
| 359 |
|
|
do { \
|
| 360 |
|
|
R##_s = X##_s ^ Y##_s; \
|
| 361 |
|
|
switch (_FP_CLS_COMBINE(X##_c, Y##_c)) \
|
| 362 |
|
|
{ \
|
| 363 |
|
|
case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_NORMAL): \
|
| 364 |
|
|
R##_c = FP_CLS_NORMAL; \
|
| 365 |
|
|
R##_e = X##_e - Y##_e; \
|
| 366 |
|
|
\
|
| 367 |
|
|
_FP_DIV_MEAT_##fs(R,X,Y); \
|
| 368 |
|
|
break; \
|
| 369 |
|
|
\
|
| 370 |
|
|
case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_NAN): \
|
| 371 |
|
|
_FP_CHOOSENAN(fs, wc, R, X, Y); \
|
| 372 |
|
|
break; \
|
| 373 |
|
|
\
|
| 374 |
|
|
case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_NORMAL): \
|
| 375 |
|
|
case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_INF): \
|
| 376 |
|
|
case _FP_CLS_COMBINE(FP_CLS_NAN,FP_CLS_ZERO): \
|
| 377 |
|
|
R##_s = X##_s; \
|
| 378 |
|
|
_FP_FRAC_COPY_##wc(R, X); \
|
| 379 |
|
|
R##_c = X##_c; \
|
| 380 |
|
|
break; \
|
| 381 |
|
|
\
|
| 382 |
|
|
case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_NAN): \
|
| 383 |
|
|
case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_NAN): \
|
| 384 |
|
|
case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_NAN): \
|
| 385 |
|
|
R##_s = Y##_s; \
|
| 386 |
|
|
_FP_FRAC_COPY_##wc(R, Y); \
|
| 387 |
|
|
R##_c = Y##_c; \
|
| 388 |
|
|
break; \
|
| 389 |
|
|
\
|
| 390 |
|
|
case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_INF): \
|
| 391 |
|
|
case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_INF): \
|
| 392 |
|
|
case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_NORMAL): \
|
| 393 |
|
|
R##_c = FP_CLS_ZERO; \
|
| 394 |
|
|
break; \
|
| 395 |
|
|
\
|
| 396 |
|
|
case _FP_CLS_COMBINE(FP_CLS_NORMAL,FP_CLS_ZERO): \
|
| 397 |
|
|
case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_ZERO): \
|
| 398 |
|
|
case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_NORMAL): \
|
| 399 |
|
|
R##_c = FP_CLS_INF; \
|
| 400 |
|
|
break; \
|
| 401 |
|
|
\
|
| 402 |
|
|
case _FP_CLS_COMBINE(FP_CLS_INF,FP_CLS_INF): \
|
| 403 |
|
|
case _FP_CLS_COMBINE(FP_CLS_ZERO,FP_CLS_ZERO): \
|
| 404 |
|
|
R##_c = FP_CLS_NAN; \
|
| 405 |
|
|
_FP_FRAC_SET_##wc(R, _FP_NANFRAC_##fs); \
|
| 406 |
|
|
break; \
|
| 407 |
|
|
\
|
| 408 |
|
|
default: \
|
| 409 |
|
|
abort(); \
|
| 410 |
|
|
} \
|
| 411 |
|
|
} while (0)
|
| 412 |
|
|
|
| 413 |
|
|
|
| 414 |
|
|
/*
|
| 415 |
|
|
* Main differential comparison routine. The inputs should be raw not
|
| 416 |
|
|
* cooked. The return is -1,0,1 for normal values, 2 otherwise.
|
| 417 |
|
|
*/
|
| 418 |
|
|
|
| 419 |
|
|
#define _FP_CMP(fs, wc, ret, X, Y, un) \
|
| 420 |
|
|
do { \
|
| 421 |
|
|
/* NANs are unordered */ \
|
| 422 |
|
|
if ((X##_e == _FP_EXPMAX_##fs && !_FP_FRAC_ZEROP_##wc(X)) \
|
| 423 |
|
|
|| (Y##_e == _FP_EXPMAX_##fs && !_FP_FRAC_ZEROP_##wc(Y))) \
|
| 424 |
|
|
{ \
|
| 425 |
|
|
ret = un; \
|
| 426 |
|
|
} \
|
| 427 |
|
|
else \
|
| 428 |
|
|
{ \
|
| 429 |
|
|
int __x_zero = (!X##_e && _FP_FRAC_ZEROP_##wc(X)) ? 1 : 0; \
|
| 430 |
|
|
int __y_zero = (!Y##_e && _FP_FRAC_ZEROP_##wc(Y)) ? 1 : 0; \
|
| 431 |
|
|
\
|
| 432 |
|
|
if (__x_zero && __y_zero) \
|
| 433 |
|
|
ret = 0; \
|
| 434 |
|
|
else if (__x_zero) \
|
| 435 |
|
|
ret = Y##_s ? 1 : -1; \
|
| 436 |
|
|
else if (__y_zero) \
|
| 437 |
|
|
ret = X##_s ? -1 : 1; \
|
| 438 |
|
|
else if (X##_s != Y##_s) \
|
| 439 |
|
|
ret = X##_s ? -1 : 1; \
|
| 440 |
|
|
else if (X##_e > Y##_e) \
|
| 441 |
|
|
ret = X##_s ? -1 : 1; \
|
| 442 |
|
|
else if (X##_e < Y##_e) \
|
| 443 |
|
|
ret = X##_s ? 1 : -1; \
|
| 444 |
|
|
else if (_FP_FRAC_GT_##wc(X, Y)) \
|
| 445 |
|
|
ret = X##_s ? -1 : 1; \
|
| 446 |
|
|
else if (_FP_FRAC_GT_##wc(Y, X)) \
|
| 447 |
|
|
ret = X##_s ? 1 : -1; \
|
| 448 |
|
|
else \
|
| 449 |
|
|
ret = 0; \
|
| 450 |
|
|
} \
|
| 451 |
|
|
} while (0)
|
| 452 |
|
|
|
| 453 |
|
|
|
| 454 |
|
|
/* Simplification for strict equality. */
|
| 455 |
|
|
|
| 456 |
|
|
#define _FP_CMP_EQ(fs, wc, ret, X, Y) \
|
| 457 |
|
|
do { \
|
| 458 |
|
|
/* NANs are unordered */ \
|
| 459 |
|
|
if ((X##_e == _FP_EXPMAX_##fs && !_FP_FRAC_ZEROP_##wc(X)) \
|
| 460 |
|
|
|| (Y##_e == _FP_EXPMAX_##fs && !_FP_FRAC_ZEROP_##wc(Y))) \
|
| 461 |
|
|
{ \
|
| 462 |
|
|
ret = 1; \
|
| 463 |
|
|
} \
|
| 464 |
|
|
else \
|
| 465 |
|
|
{ \
|
| 466 |
|
|
ret = !(X##_e == Y##_e \
|
| 467 |
|
|
&& _FP_FRAC_EQ_##wc(X, Y) \
|
| 468 |
|
|
&& (X##_s == Y##_s || !X##_e && _FP_FRAC_ZEROP_##wc(X))); \
|
| 469 |
|
|
} \
|
| 470 |
|
|
} while (0)
|
| 471 |
|
|
|
| 472 |
|
|
/*
|
| 473 |
|
|
* Main square root routine. The input value should be cooked.
|
| 474 |
|
|
*/
|
| 475 |
|
|
|
| 476 |
|
|
#define _FP_SQRT(fs, wc, R, X) \
|
| 477 |
|
|
do { \
|
| 478 |
|
|
_FP_FRAC_DECL_##wc(T); _FP_FRAC_DECL_##wc(S); \
|
| 479 |
|
|
_FP_W_TYPE q; \
|
| 480 |
|
|
switch (X##_c) \
|
| 481 |
|
|
{ \
|
| 482 |
|
|
case FP_CLS_NAN: \
|
| 483 |
|
|
R##_s = 0; \
|
| 484 |
|
|
R##_c = FP_CLS_NAN; \
|
| 485 |
|
|
_FP_FRAC_SET_##wc(X, _FP_ZEROFRAC_##wc); \
|
| 486 |
|
|
break; \
|
| 487 |
|
|
case FP_CLS_INF: \
|
| 488 |
|
|
if (X##_s) \
|
| 489 |
|
|
{ \
|
| 490 |
|
|
R##_s = 0; \
|
| 491 |
|
|
R##_c = FP_CLS_NAN; /* sNAN */ \
|
| 492 |
|
|
} \
|
| 493 |
|
|
else \
|
| 494 |
|
|
{ \
|
| 495 |
|
|
R##_s = 0; \
|
| 496 |
|
|
R##_c = FP_CLS_INF; /* sqrt(+inf) = +inf */ \
|
| 497 |
|
|
} \
|
| 498 |
|
|
break; \
|
| 499 |
|
|
case FP_CLS_ZERO: \
|
| 500 |
|
|
R##_s = X##_s; \
|
| 501 |
|
|
R##_c = FP_CLS_ZERO; /* sqrt(+-0) = +-0 */ \
|
| 502 |
|
|
break; \
|
| 503 |
|
|
case FP_CLS_NORMAL: \
|
| 504 |
|
|
R##_s = 0; \
|
| 505 |
|
|
if (X##_s) \
|
| 506 |
|
|
{ \
|
| 507 |
|
|
R##_c = FP_CLS_NAN; /* sNAN */ \
|
| 508 |
|
|
break; \
|
| 509 |
|
|
} \
|
| 510 |
|
|
R##_c = FP_CLS_NORMAL; \
|
| 511 |
|
|
if (X##_e & 1) \
|
| 512 |
|
|
_FP_FRAC_SLL_##wc(X, 1); \
|
| 513 |
|
|
R##_e = X##_e >> 1; \
|
| 514 |
|
|
_FP_FRAC_SET_##wc(S, _FP_ZEROFRAC_##wc); \
|
| 515 |
|
|
_FP_FRAC_SET_##wc(R, _FP_ZEROFRAC_##wc); \
|
| 516 |
|
|
q = _FP_OVERFLOW_##fs; \
|
| 517 |
|
|
_FP_FRAC_SLL_##wc(X, 1); \
|
| 518 |
|
|
_FP_SQRT_MEAT_##wc(R, S, T, X, q); \
|
| 519 |
|
|
_FP_FRAC_SRL_##wc(R, 1); \
|
| 520 |
|
|
} \
|
| 521 |
|
|
} while (0)
|
| 522 |
|
|
|
| 523 |
|
|
/*
|
| 524 |
|
|
* Convert from FP to integer
|
| 525 |
|
|
*/
|
| 526 |
|
|
|
| 527 |
|
|
/* "When a NaN, infinity, large positive argument >= 2147483648.0, or
|
| 528 |
|
|
* large negative argument <= -2147483649.0 is converted to an integer,
|
| 529 |
|
|
* the invalid_current bit...should be set and fp_exception_IEEE_754 should
|
| 530 |
|
|
* be raised. If the floating point invalid trap is disabled, no trap occurs
|
| 531 |
|
|
* and a numerical result is generated: if the sign bit of the operand
|
| 532 |
|
|
* is 0, the result is 2147483647; if the sign bit of the operand is 1,
|
| 533 |
|
|
* the result is -2147483648."
|
| 534 |
|
|
* Similarly for conversion to extended ints, except that the boundaries
|
| 535 |
|
|
* are >= 2^63, <= -(2^63 + 1), and the results are 2^63 + 1 for s=0 and
|
| 536 |
|
|
* -2^63 for s=1.
|
| 537 |
|
|
* -- SPARC Architecture Manual V9, Appendix B, which specifies how
|
| 538 |
|
|
* SPARCs resolve implementation dependencies in the IEEE-754 spec.
|
| 539 |
|
|
* I don't believe that the code below follows this. I'm not even sure
|
| 540 |
|
|
* it's right!
|
| 541 |
|
|
* It doesn't cope with needing to convert to an n bit integer when there
|
| 542 |
|
|
* is no n bit integer type. Fortunately gcc provides long long so this
|
| 543 |
|
|
* isn't a problem for sparc32.
|
| 544 |
|
|
* I have, however, fixed its NaN handling to conform as above.
|
| 545 |
|
|
* -- PMM 02/1998
|
| 546 |
|
|
* NB: rsigned is not 'is r declared signed?' but 'should the value stored
|
| 547 |
|
|
* in r be signed or unsigned?'. r is always(?) declared unsigned.
|
| 548 |
|
|
* Comments below are mine, BTW -- PMM
|
| 549 |
|
|
*/
|
| 550 |
|
|
#define _FP_TO_INT(fs, wc, r, X, rsize, rsigned) \
|
| 551 |
|
|
do { \
|
| 552 |
|
|
switch (X##_c) \
|
| 553 |
|
|
{ \
|
| 554 |
|
|
case FP_CLS_NORMAL: \
|
| 555 |
|
|
if (X##_e < 0) \
|
| 556 |
|
|
{ \
|
| 557 |
|
|
/* case FP_CLS_NAN: see above! */ \
|
| 558 |
|
|
case FP_CLS_ZERO: \
|
| 559 |
|
|
r = 0; \
|
| 560 |
|
|
} \
|
| 561 |
|
|
else if (X##_e >= rsize - (rsigned != 0)) \
|
| 562 |
|
|
{ /* overflow */ \
|
| 563 |
|
|
case FP_CLS_NAN: \
|
| 564 |
|
|
case FP_CLS_INF: \
|
| 565 |
|
|
if (rsigned) \
|
| 566 |
|
|
{ \
|
| 567 |
|
|
r = 1; \
|
| 568 |
|
|
r <<= rsize - 1; \
|
| 569 |
|
|
r -= 1 - X##_s; \
|
| 570 |
|
|
} \
|
| 571 |
|
|
else \
|
| 572 |
|
|
{ \
|
| 573 |
|
|
r = 0; \
|
| 574 |
|
|
if (!X##_s) \
|
| 575 |
|
|
r = ~r; \
|
| 576 |
|
|
} \
|
| 577 |
|
|
} \
|
| 578 |
|
|
else \
|
| 579 |
|
|
{ \
|
| 580 |
|
|
if (_FP_W_TYPE_SIZE*wc < rsize) \
|
| 581 |
|
|
{ \
|
| 582 |
|
|
_FP_FRAC_ASSEMBLE_##wc(r, X, rsize); \
|
| 583 |
|
|
r <<= X##_e - _FP_WFRACBITS_##fs; \
|
| 584 |
|
|
} \
|
| 585 |
|
|
else \
|
| 586 |
|
|
{ \
|
| 587 |
|
|
if (X##_e >= _FP_WFRACBITS_##fs) \
|
| 588 |
|
|
_FP_FRAC_SLL_##wc(X, (X##_e - _FP_WFRACBITS_##fs + 1));\
|
| 589 |
|
|
else \
|
| 590 |
|
|
_FP_FRAC_SRL_##wc(X, (_FP_WFRACBITS_##fs - X##_e - 1));\
|
| 591 |
|
|
_FP_FRAC_ASSEMBLE_##wc(r, X, rsize); \
|
| 592 |
|
|
} \
|
| 593 |
|
|
if (rsigned && X##_s) \
|
| 594 |
|
|
r = -r; \
|
| 595 |
|
|
} \
|
| 596 |
|
|
break; \
|
| 597 |
|
|
} \
|
| 598 |
|
|
} while (0)
|
| 599 |
|
|
|
| 600 |
|
|
#define _FP_FROM_INT(fs, wc, X, r, rsize, rtype) \
|
| 601 |
|
|
do { \
|
| 602 |
|
|
if (r) \
|
| 603 |
|
|
{ \
|
| 604 |
|
|
X##_c = FP_CLS_NORMAL; \
|
| 605 |
|
|
\
|
| 606 |
|
|
if ((X##_s = (r < 0))) \
|
| 607 |
|
|
r = -r; \
|
| 608 |
|
|
/* Note that `r' is now considered unsigned, so we don't have \
|
| 609 |
|
|
to worry about the single signed overflow case. */ \
|
| 610 |
|
|
\
|
| 611 |
|
|
if (rsize <= _FP_W_TYPE_SIZE) \
|
| 612 |
|
|
__FP_CLZ(X##_e, r); \
|
| 613 |
|
|
else \
|
| 614 |
|
|
__FP_CLZ_2(X##_e, (_FP_W_TYPE)(r >> _FP_W_TYPE_SIZE), \
|
| 615 |
|
|
(_FP_W_TYPE)r); \
|
| 616 |
|
|
if (rsize < _FP_W_TYPE_SIZE) \
|
| 617 |
|
|
X##_e -= (_FP_W_TYPE_SIZE - rsize); \
|
| 618 |
|
|
X##_e = rsize - X##_e - 1; \
|
| 619 |
|
|
\
|
| 620 |
|
|
if (_FP_FRACBITS_##fs < rsize && _FP_WFRACBITS_##fs < X##_e) \
|
| 621 |
|
|
__FP_FRAC_SRS_1(r, (X##_e - _FP_WFRACBITS_##fs), rsize); \
|
| 622 |
|
|
r &= ~((_FP_W_TYPE)1 << X##_e); \
|
| 623 |
|
|
_FP_FRAC_DISASSEMBLE_##wc(X, ((unsigned rtype)r), rsize); \
|
| 624 |
|
|
_FP_FRAC_SLL_##wc(X, (_FP_WFRACBITS_##fs - X##_e - 1)); \
|
| 625 |
|
|
} \
|
| 626 |
|
|
else \
|
| 627 |
|
|
{ \
|
| 628 |
|
|
X##_c = FP_CLS_ZERO, X##_s = 0; \
|
| 629 |
|
|
} \
|
| 630 |
|
|
} while (0)
|
| 631 |
|
|
|
| 632 |
|
|
|
| 633 |
|
|
#define FP_CONV(dfs,sfs,dwc,swc,D,S) \
|
| 634 |
|
|
do { \
|
| 635 |
|
|
_FP_FRAC_CONV_##dwc##_##swc(dfs, sfs, D, S); \
|
| 636 |
|
|
D##_e = S##_e; \
|
| 637 |
|
|
D##_c = S##_c; \
|
| 638 |
|
|
D##_s = S##_s; \
|
| 639 |
|
|
} while (0)
|
| 640 |
|
|
|
| 641 |
|
|
/*
|
| 642 |
|
|
* Helper primitives.
|
| 643 |
|
|
*/
|
| 644 |
|
|
|
| 645 |
|
|
/* Count leading zeros in a word. */
|
| 646 |
|
|
|
| 647 |
|
|
#ifndef __FP_CLZ
|
| 648 |
|
|
#if _FP_W_TYPE_SIZE < 64
|
| 649 |
|
|
/* this is just to shut the compiler up about shifts > word length -- PMM 02/1998 */
|
| 650 |
|
|
#define __FP_CLZ(r, x) \
|
| 651 |
|
|
do { \
|
| 652 |
|
|
_FP_W_TYPE _t = (x); \
|
| 653 |
|
|
r = _FP_W_TYPE_SIZE - 1; \
|
| 654 |
|
|
if (_t > 0xffff) r -= 16; \
|
| 655 |
|
|
if (_t > 0xffff) _t >>= 16; \
|
| 656 |
|
|
if (_t > 0xff) r -= 8; \
|
| 657 |
|
|
if (_t > 0xff) _t >>= 8; \
|
| 658 |
|
|
if (_t & 0xf0) r -= 4; \
|
| 659 |
|
|
if (_t & 0xf0) _t >>= 4; \
|
| 660 |
|
|
if (_t & 0xc) r -= 2; \
|
| 661 |
|
|
if (_t & 0xc) _t >>= 2; \
|
| 662 |
|
|
if (_t & 0x2) r -= 1; \
|
| 663 |
|
|
} while (0)
|
| 664 |
|
|
#else /* not _FP_W_TYPE_SIZE < 64 */
|
| 665 |
|
|
#define __FP_CLZ(r, x) \
|
| 666 |
|
|
do { \
|
| 667 |
|
|
_FP_W_TYPE _t = (x); \
|
| 668 |
|
|
r = _FP_W_TYPE_SIZE - 1; \
|
| 669 |
|
|
if (_t > 0xffffffff) r -= 32; \
|
| 670 |
|
|
if (_t > 0xffffffff) _t >>= 32; \
|
| 671 |
|
|
if (_t > 0xffff) r -= 16; \
|
| 672 |
|
|
if (_t > 0xffff) _t >>= 16; \
|
| 673 |
|
|
if (_t > 0xff) r -= 8; \
|
| 674 |
|
|
if (_t > 0xff) _t >>= 8; \
|
| 675 |
|
|
if (_t & 0xf0) r -= 4; \
|
| 676 |
|
|
if (_t & 0xf0) _t >>= 4; \
|
| 677 |
|
|
if (_t & 0xc) r -= 2; \
|
| 678 |
|
|
if (_t & 0xc) _t >>= 2; \
|
| 679 |
|
|
if (_t & 0x2) r -= 1; \
|
| 680 |
|
|
} while (0)
|
| 681 |
|
|
#endif /* not _FP_W_TYPE_SIZE < 64 */
|
| 682 |
|
|
#endif /* ndef __FP_CLZ */
|
| 683 |
|
|
|
| 684 |
|
|
#define _FP_DIV_HELP_imm(q, r, n, d) \
|
| 685 |
|
|
do { \
|
| 686 |
|
|
q = n / d, r = n % d; \
|
| 687 |
|
|
} while (0)
|
| 688 |
|
|
|
