George Lander | da55ef9 | 2015-11-19 12:05:06 +0000 | [diff] [blame] | 1 | /* |
Szabolcs Nagy | 1b94597 | 2018-05-14 14:46:40 +0100 | [diff] [blame] | 2 | * semi.c: test implementations of mathlib seminumerical functions |
George Lander | da55ef9 | 2015-11-19 12:05:06 +0000 | [diff] [blame] | 3 | * |
Szabolcs Nagy | 11253b0 | 2018-11-12 11:10:57 +0000 | [diff] [blame] | 4 | * Copyright (c) 1999-2018, Arm Limited. |
| 5 | * SPDX-License-Identifier: MIT |
George Lander | da55ef9 | 2015-11-19 12:05:06 +0000 | [diff] [blame] | 6 | */ |
| 7 | |
| 8 | #include <stdio.h> |
| 9 | #include "semi.h" |
| 10 | |
| 11 | static void test_rint(uint32 *in, uint32 *out, |
| 12 | int isfloor, int isceil) { |
| 13 | int sign = in[0] & 0x80000000; |
| 14 | int roundup = (isfloor && sign) || (isceil && !sign); |
| 15 | uint32 xh, xl, roundword; |
| 16 | int ex = (in[0] >> 20) & 0x7FF; /* exponent */ |
| 17 | int i; |
| 18 | |
| 19 | if ((ex > 0x3ff + 52 - 1) || /* things this big can't be fractional */ |
| 20 | ((in[0] & 0x7FFFFFFF) == 0 && in[1] == 0)) { /* zero */ |
| 21 | /* NaN, Inf, a large integer, or zero: just return the input */ |
| 22 | out[0] = in[0]; |
| 23 | out[1] = in[1]; |
| 24 | return; |
| 25 | } |
| 26 | |
| 27 | /* |
| 28 | * Special case: ex < 0x3ff, ie our number is in (0,1). Return |
| 29 | * 1 or 0 according to roundup. |
| 30 | */ |
| 31 | if (ex < 0x3ff) { |
| 32 | out[0] = sign | (roundup ? 0x3FF00000 : 0); |
| 33 | out[1] = 0; |
| 34 | return; |
| 35 | } |
| 36 | |
| 37 | /* |
| 38 | * We're not short of time here, so we'll do this the hideously |
| 39 | * inefficient way. Shift bit by bit so that the units place is |
| 40 | * somewhere predictable, round, and shift back again. |
| 41 | */ |
| 42 | xh = in[0]; |
| 43 | xl = in[1]; |
| 44 | roundword = 0; |
| 45 | for (i = ex; i < 0x3ff + 52; i++) { |
| 46 | if (roundword & 1) |
| 47 | roundword |= 2; /* preserve sticky bit */ |
| 48 | roundword = (roundword >> 1) | ((xl & 1) << 31); |
| 49 | xl = (xl >> 1) | ((xh & 1) << 31); |
| 50 | xh = xh >> 1; |
| 51 | } |
| 52 | if (roundword && roundup) { |
| 53 | xl++; |
| 54 | xh += (xl==0); |
| 55 | } |
| 56 | for (i = ex; i < 0x3ff + 52; i++) { |
| 57 | xh = (xh << 1) | ((xl >> 31) & 1); |
| 58 | xl = (xl & 0x7FFFFFFF) << 1; |
| 59 | } |
| 60 | out[0] = xh; |
| 61 | out[1] = xl; |
| 62 | } |
| 63 | |
| 64 | char *test_ceil(uint32 *in, uint32 *out) { |
| 65 | test_rint(in, out, 0, 1); |
| 66 | return NULL; |
| 67 | } |
| 68 | |
| 69 | char *test_floor(uint32 *in, uint32 *out) { |
| 70 | test_rint(in, out, 1, 0); |
| 71 | return NULL; |
| 72 | } |
| 73 | |
| 74 | static void test_rintf(uint32 *in, uint32 *out, |
| 75 | int isfloor, int isceil) { |
| 76 | int sign = *in & 0x80000000; |
| 77 | int roundup = (isfloor && sign) || (isceil && !sign); |
| 78 | uint32 x, roundword; |
| 79 | int ex = (*in >> 23) & 0xFF; /* exponent */ |
| 80 | int i; |
| 81 | |
| 82 | if ((ex > 0x7f + 23 - 1) || /* things this big can't be fractional */ |
| 83 | (*in & 0x7FFFFFFF) == 0) { /* zero */ |
| 84 | /* NaN, Inf, a large integer, or zero: just return the input */ |
| 85 | *out = *in; |
| 86 | return; |
| 87 | } |
| 88 | |
| 89 | /* |
| 90 | * Special case: ex < 0x7f, ie our number is in (0,1). Return |
| 91 | * 1 or 0 according to roundup. |
| 92 | */ |
| 93 | if (ex < 0x7f) { |
| 94 | *out = sign | (roundup ? 0x3F800000 : 0); |
| 95 | return; |
| 96 | } |
| 97 | |
| 98 | /* |
| 99 | * We're not short of time here, so we'll do this the hideously |
| 100 | * inefficient way. Shift bit by bit so that the units place is |
| 101 | * somewhere predictable, round, and shift back again. |
| 102 | */ |
| 103 | x = *in; |
| 104 | roundword = 0; |
| 105 | for (i = ex; i < 0x7F + 23; i++) { |
| 106 | if (roundword & 1) |
| 107 | roundword |= 2; /* preserve sticky bit */ |
| 108 | roundword = (roundword >> 1) | ((x & 1) << 31); |
| 109 | x = x >> 1; |
| 110 | } |
| 111 | if (roundword && roundup) { |
| 112 | x++; |
| 113 | } |
| 114 | for (i = ex; i < 0x7F + 23; i++) { |
| 115 | x = x << 1; |
| 116 | } |
| 117 | *out = x; |
| 118 | } |
| 119 | |
| 120 | char *test_ceilf(uint32 *in, uint32 *out) { |
| 121 | test_rintf(in, out, 0, 1); |
| 122 | return NULL; |
| 123 | } |
| 124 | |
| 125 | char *test_floorf(uint32 *in, uint32 *out) { |
| 126 | test_rintf(in, out, 1, 0); |
| 127 | return NULL; |
| 128 | } |
| 129 | |
| 130 | char *test_fmod(uint32 *a, uint32 *b, uint32 *out) { |
| 131 | int sign; |
| 132 | int32 aex, bex; |
| 133 | uint32 am[2], bm[2]; |
| 134 | |
| 135 | if (((a[0] & 0x7FFFFFFF) << 1) + !!a[1] > 0xFFE00000 || |
| 136 | ((b[0] & 0x7FFFFFFF) << 1) + !!b[1] > 0xFFE00000) { |
| 137 | /* a or b is NaN: return QNaN, optionally with IVO */ |
| 138 | uint32 an, bn; |
| 139 | out[0] = 0x7ff80000; |
| 140 | out[1] = 1; |
| 141 | an = ((a[0] & 0x7FFFFFFF) << 1) + !!a[1]; |
| 142 | bn = ((b[0] & 0x7FFFFFFF) << 1) + !!b[1]; |
| 143 | if ((an > 0xFFE00000 && an < 0xFFF00000) || |
| 144 | (bn > 0xFFE00000 && bn < 0xFFF00000)) |
| 145 | return "i"; /* at least one SNaN: IVO */ |
| 146 | else |
| 147 | return NULL; /* no SNaNs, but at least 1 QNaN */ |
| 148 | } |
| 149 | if ((b[0] & 0x7FFFFFFF) == 0 && b[1] == 0) { /* b==0: EDOM */ |
| 150 | out[0] = 0x7ff80000; |
| 151 | out[1] = 1; |
| 152 | return "EDOM status=i"; |
| 153 | } |
| 154 | if ((a[0] & 0x7FF00000) == 0x7FF00000) { /* a==Inf: EDOM */ |
| 155 | out[0] = 0x7ff80000; |
| 156 | out[1] = 1; |
| 157 | return "EDOM status=i"; |
| 158 | } |
| 159 | if ((b[0] & 0x7FF00000) == 0x7FF00000) { /* b==Inf: return a */ |
| 160 | out[0] = a[0]; |
| 161 | out[1] = a[1]; |
| 162 | return NULL; |
| 163 | } |
| 164 | if ((a[0] & 0x7FFFFFFF) == 0 && a[1] == 0) { /* a==0: return a */ |
| 165 | out[0] = a[0]; |
| 166 | out[1] = a[1]; |
| 167 | return NULL; |
| 168 | } |
| 169 | |
| 170 | /* |
| 171 | * OK. That's the special cases cleared out of the way. Now we |
| 172 | * have finite (though not necessarily normal) a and b. |
| 173 | */ |
| 174 | sign = a[0] & 0x80000000; /* we discard sign of b */ |
| 175 | test_frexp(a, am, (uint32 *)&aex); |
| 176 | test_frexp(b, bm, (uint32 *)&bex); |
| 177 | am[0] &= 0xFFFFF, am[0] |= 0x100000; |
| 178 | bm[0] &= 0xFFFFF, bm[0] |= 0x100000; |
| 179 | |
| 180 | while (aex >= bex) { |
| 181 | if (am[0] > bm[0] || (am[0] == bm[0] && am[1] >= bm[1])) { |
| 182 | am[1] -= bm[1]; |
| 183 | am[0] = am[0] - bm[0] - (am[1] > ~bm[1]); |
| 184 | } |
| 185 | if (aex > bex) { |
| 186 | am[0] = (am[0] << 1) | ((am[1] & 0x80000000) >> 31); |
| 187 | am[1] <<= 1; |
| 188 | aex--; |
| 189 | } else |
| 190 | break; |
| 191 | } |
| 192 | |
| 193 | /* |
| 194 | * Renormalise final result; this can be cunningly done by |
| 195 | * passing a denormal to ldexp. |
| 196 | */ |
| 197 | aex += 0x3fd; |
| 198 | am[0] |= sign; |
| 199 | test_ldexp(am, (uint32 *)&aex, out); |
| 200 | |
| 201 | return NULL; /* FIXME */ |
| 202 | } |
| 203 | |
| 204 | char *test_fmodf(uint32 *a, uint32 *b, uint32 *out) { |
| 205 | int sign; |
| 206 | int32 aex, bex; |
| 207 | uint32 am, bm; |
| 208 | |
| 209 | if ((*a & 0x7FFFFFFF) > 0x7F800000 || |
| 210 | (*b & 0x7FFFFFFF) > 0x7F800000) { |
| 211 | /* a or b is NaN: return QNaN, optionally with IVO */ |
| 212 | uint32 an, bn; |
| 213 | *out = 0x7fc00001; |
| 214 | an = *a & 0x7FFFFFFF; |
| 215 | bn = *b & 0x7FFFFFFF; |
| 216 | if ((an > 0x7f800000 && an < 0x7fc00000) || |
| 217 | (bn > 0x7f800000 && bn < 0x7fc00000)) |
| 218 | return "i"; /* at least one SNaN: IVO */ |
| 219 | else |
| 220 | return NULL; /* no SNaNs, but at least 1 QNaN */ |
| 221 | } |
| 222 | if ((*b & 0x7FFFFFFF) == 0) { /* b==0: EDOM */ |
| 223 | *out = 0x7fc00001; |
| 224 | return "EDOM status=i"; |
| 225 | } |
| 226 | if ((*a & 0x7F800000) == 0x7F800000) { /* a==Inf: EDOM */ |
| 227 | *out = 0x7fc00001; |
| 228 | return "EDOM status=i"; |
| 229 | } |
| 230 | if ((*b & 0x7F800000) == 0x7F800000) { /* b==Inf: return a */ |
| 231 | *out = *a; |
| 232 | return NULL; |
| 233 | } |
| 234 | if ((*a & 0x7FFFFFFF) == 0) { /* a==0: return a */ |
| 235 | *out = *a; |
| 236 | return NULL; |
| 237 | } |
| 238 | |
| 239 | /* |
| 240 | * OK. That's the special cases cleared out of the way. Now we |
| 241 | * have finite (though not necessarily normal) a and b. |
| 242 | */ |
| 243 | sign = a[0] & 0x80000000; /* we discard sign of b */ |
| 244 | test_frexpf(a, &am, (uint32 *)&aex); |
| 245 | test_frexpf(b, &bm, (uint32 *)&bex); |
| 246 | am &= 0x7FFFFF, am |= 0x800000; |
| 247 | bm &= 0x7FFFFF, bm |= 0x800000; |
| 248 | |
| 249 | while (aex >= bex) { |
| 250 | if (am >= bm) { |
| 251 | am -= bm; |
| 252 | } |
| 253 | if (aex > bex) { |
| 254 | am <<= 1; |
| 255 | aex--; |
| 256 | } else |
| 257 | break; |
| 258 | } |
| 259 | |
| 260 | /* |
| 261 | * Renormalise final result; this can be cunningly done by |
| 262 | * passing a denormal to ldexp. |
| 263 | */ |
| 264 | aex += 0x7d; |
| 265 | am |= sign; |
| 266 | test_ldexpf(&am, (uint32 *)&aex, out); |
| 267 | |
| 268 | return NULL; /* FIXME */ |
| 269 | } |
| 270 | |
| 271 | char *test_ldexp(uint32 *x, uint32 *np, uint32 *out) { |
| 272 | int n = *np; |
| 273 | int32 n2; |
| 274 | uint32 y[2]; |
| 275 | int ex = (x[0] >> 20) & 0x7FF; /* exponent */ |
| 276 | int sign = x[0] & 0x80000000; |
| 277 | |
| 278 | if (ex == 0x7FF) { /* inf/NaN; just return x */ |
| 279 | out[0] = x[0]; |
| 280 | out[1] = x[1]; |
| 281 | return NULL; |
| 282 | } |
| 283 | if ((x[0] & 0x7FFFFFFF) == 0 && x[1] == 0) { /* zero: return x */ |
| 284 | out[0] = x[0]; |
| 285 | out[1] = x[1]; |
| 286 | return NULL; |
| 287 | } |
| 288 | |
| 289 | test_frexp(x, y, (uint32 *)&n2); |
| 290 | ex = n + n2; |
| 291 | if (ex > 0x400) { /* overflow */ |
| 292 | out[0] = sign | 0x7FF00000; |
| 293 | out[1] = 0; |
| 294 | return "overflow"; |
| 295 | } |
| 296 | /* |
| 297 | * Underflow. 2^-1074 is 00000000.00000001; so if ex == -1074 |
| 298 | * then we have something [2^-1075,2^-1074). Under round-to- |
| 299 | * nearest-even, this whole interval rounds up to 2^-1074, |
| 300 | * except for the bottom endpoint which rounds to even and is |
| 301 | * an underflow condition. |
| 302 | * |
| 303 | * So, ex < -1074 is definite underflow, and ex == -1074 is |
| 304 | * underflow iff all mantissa bits are zero. |
| 305 | */ |
| 306 | if (ex < -1074 || (ex == -1074 && (y[0] & 0xFFFFF) == 0 && y[1] == 0)) { |
| 307 | out[0] = sign; /* underflow: correctly signed zero */ |
| 308 | out[1] = 0; |
| 309 | return "underflow"; |
| 310 | } |
| 311 | |
| 312 | /* |
| 313 | * No overflow or underflow; should be nice and simple, unless |
| 314 | * we have to denormalise and round the result. |
| 315 | */ |
| 316 | if (ex < -1021) { /* denormalise and round */ |
| 317 | uint32 roundword; |
| 318 | y[0] &= 0x000FFFFF; |
| 319 | y[0] |= 0x00100000; /* set leading bit */ |
| 320 | roundword = 0; |
| 321 | while (ex < -1021) { |
| 322 | if (roundword & 1) |
| 323 | roundword |= 2; /* preserve sticky bit */ |
| 324 | roundword = (roundword >> 1) | ((y[1] & 1) << 31); |
| 325 | y[1] = (y[1] >> 1) | ((y[0] & 1) << 31); |
| 326 | y[0] = y[0] >> 1; |
| 327 | ex++; |
| 328 | } |
| 329 | if (roundword > 0x80000000 || /* round up */ |
| 330 | (roundword == 0x80000000 && (y[1] & 1))) { /* round up to even */ |
| 331 | y[1]++; |
| 332 | y[0] += (y[1] == 0); |
| 333 | } |
| 334 | out[0] = sign | y[0]; |
| 335 | out[1] = y[1]; |
| 336 | /* Proper ERANGE underflow was handled earlier, but we still |
| 337 | * expect an IEEE Underflow exception if this partially |
| 338 | * underflowed result is not exact. */ |
| 339 | if (roundword) |
| 340 | return "u"; |
| 341 | return NULL; /* underflow was handled earlier */ |
| 342 | } else { |
| 343 | out[0] = y[0] + (ex << 20); |
| 344 | out[1] = y[1]; |
| 345 | return NULL; |
| 346 | } |
| 347 | } |
| 348 | |
| 349 | char *test_ldexpf(uint32 *x, uint32 *np, uint32 *out) { |
| 350 | int n = *np; |
| 351 | int32 n2; |
| 352 | uint32 y; |
| 353 | int ex = (*x >> 23) & 0xFF; /* exponent */ |
| 354 | int sign = *x & 0x80000000; |
| 355 | |
| 356 | if (ex == 0xFF) { /* inf/NaN; just return x */ |
| 357 | *out = *x; |
| 358 | return NULL; |
| 359 | } |
| 360 | if ((*x & 0x7FFFFFFF) == 0) { /* zero: return x */ |
| 361 | *out = *x; |
| 362 | return NULL; |
| 363 | } |
| 364 | |
| 365 | test_frexpf(x, &y, (uint32 *)&n2); |
| 366 | ex = n + n2; |
| 367 | if (ex > 0x80) { /* overflow */ |
| 368 | *out = sign | 0x7F800000; |
| 369 | return "overflow"; |
| 370 | } |
| 371 | /* |
| 372 | * Underflow. 2^-149 is 00000001; so if ex == -149 then we have |
| 373 | * something [2^-150,2^-149). Under round-to- nearest-even, |
| 374 | * this whole interval rounds up to 2^-149, except for the |
| 375 | * bottom endpoint which rounds to even and is an underflow |
| 376 | * condition. |
| 377 | * |
| 378 | * So, ex < -149 is definite underflow, and ex == -149 is |
| 379 | * underflow iff all mantissa bits are zero. |
| 380 | */ |
| 381 | if (ex < -149 || (ex == -149 && (y & 0x7FFFFF) == 0)) { |
| 382 | *out = sign; /* underflow: correctly signed zero */ |
| 383 | return "underflow"; |
| 384 | } |
| 385 | |
| 386 | /* |
| 387 | * No overflow or underflow; should be nice and simple, unless |
| 388 | * we have to denormalise and round the result. |
| 389 | */ |
| 390 | if (ex < -125) { /* denormalise and round */ |
| 391 | uint32 roundword; |
| 392 | y &= 0x007FFFFF; |
| 393 | y |= 0x00800000; /* set leading bit */ |
| 394 | roundword = 0; |
| 395 | while (ex < -125) { |
| 396 | if (roundword & 1) |
| 397 | roundword |= 2; /* preserve sticky bit */ |
| 398 | roundword = (roundword >> 1) | ((y & 1) << 31); |
| 399 | y = y >> 1; |
| 400 | ex++; |
| 401 | } |
| 402 | if (roundword > 0x80000000 || /* round up */ |
| 403 | (roundword == 0x80000000 && (y & 1))) { /* round up to even */ |
| 404 | y++; |
| 405 | } |
| 406 | *out = sign | y; |
| 407 | /* Proper ERANGE underflow was handled earlier, but we still |
| 408 | * expect an IEEE Underflow exception if this partially |
| 409 | * underflowed result is not exact. */ |
| 410 | if (roundword) |
| 411 | return "u"; |
| 412 | return NULL; /* underflow was handled earlier */ |
| 413 | } else { |
| 414 | *out = y + (ex << 23); |
| 415 | return NULL; |
| 416 | } |
| 417 | } |
| 418 | |
| 419 | char *test_frexp(uint32 *x, uint32 *out, uint32 *nout) { |
| 420 | int ex = (x[0] >> 20) & 0x7FF; /* exponent */ |
| 421 | if (ex == 0x7FF) { /* inf/NaN; return x/0 */ |
| 422 | out[0] = x[0]; |
| 423 | out[1] = x[1]; |
| 424 | nout[0] = 0; |
| 425 | return NULL; |
| 426 | } |
| 427 | if (ex == 0) { /* denormals/zeros */ |
| 428 | int sign; |
| 429 | uint32 xh, xl; |
| 430 | if ((x[0] & 0x7FFFFFFF) == 0 && x[1] == 0) { |
| 431 | /* zero: return x/0 */ |
| 432 | out[0] = x[0]; |
| 433 | out[1] = x[1]; |
| 434 | nout[0] = 0; |
| 435 | return NULL; |
| 436 | } |
| 437 | sign = x[0] & 0x80000000; |
| 438 | xh = x[0] & 0x7FFFFFFF; |
| 439 | xl = x[1]; |
| 440 | ex = 1; |
| 441 | while (!(xh & 0x100000)) { |
| 442 | ex--; |
| 443 | xh = (xh << 1) | ((xl >> 31) & 1); |
| 444 | xl = (xl & 0x7FFFFFFF) << 1; |
| 445 | } |
| 446 | out[0] = sign | 0x3FE00000 | (xh & 0xFFFFF); |
| 447 | out[1] = xl; |
| 448 | nout[0] = ex - 0x3FE; |
| 449 | return NULL; |
| 450 | } |
| 451 | out[0] = 0x3FE00000 | (x[0] & 0x800FFFFF); |
| 452 | out[1] = x[1]; |
| 453 | nout[0] = ex - 0x3FE; |
| 454 | return NULL; /* ordinary number; no error */ |
| 455 | } |
| 456 | |
| 457 | char *test_frexpf(uint32 *x, uint32 *out, uint32 *nout) { |
| 458 | int ex = (*x >> 23) & 0xFF; /* exponent */ |
| 459 | if (ex == 0xFF) { /* inf/NaN; return x/0 */ |
| 460 | *out = *x; |
| 461 | nout[0] = 0; |
| 462 | return NULL; |
| 463 | } |
| 464 | if (ex == 0) { /* denormals/zeros */ |
| 465 | int sign; |
| 466 | uint32 xv; |
| 467 | if ((*x & 0x7FFFFFFF) == 0) { |
| 468 | /* zero: return x/0 */ |
| 469 | *out = *x; |
| 470 | nout[0] = 0; |
| 471 | return NULL; |
| 472 | } |
| 473 | sign = *x & 0x80000000; |
| 474 | xv = *x & 0x7FFFFFFF; |
| 475 | ex = 1; |
| 476 | while (!(xv & 0x800000)) { |
| 477 | ex--; |
| 478 | xv = xv << 1; |
| 479 | } |
| 480 | *out = sign | 0x3F000000 | (xv & 0x7FFFFF); |
| 481 | nout[0] = ex - 0x7E; |
| 482 | return NULL; |
| 483 | } |
| 484 | *out = 0x3F000000 | (*x & 0x807FFFFF); |
| 485 | nout[0] = ex - 0x7E; |
| 486 | return NULL; /* ordinary number; no error */ |
| 487 | } |
| 488 | |
| 489 | char *test_modf(uint32 *x, uint32 *fout, uint32 *iout) { |
| 490 | int ex = (x[0] >> 20) & 0x7FF; /* exponent */ |
| 491 | int sign = x[0] & 0x80000000; |
| 492 | uint32 fh, fl; |
| 493 | |
| 494 | if (((x[0] & 0x7FFFFFFF) | (!!x[1])) > 0x7FF00000) { |
| 495 | /* |
| 496 | * NaN input: return the same in _both_ outputs. |
| 497 | */ |
| 498 | fout[0] = iout[0] = x[0]; |
| 499 | fout[1] = iout[1] = x[1]; |
| 500 | return NULL; |
| 501 | } |
| 502 | |
| 503 | test_rint(x, iout, 0, 0); |
| 504 | fh = x[0] - iout[0]; |
| 505 | fl = x[1] - iout[1]; |
| 506 | if (!fh && !fl) { /* no fraction part */ |
| 507 | fout[0] = sign; |
| 508 | fout[1] = 0; |
| 509 | return NULL; |
| 510 | } |
| 511 | if (!(iout[0] & 0x7FFFFFFF) && !iout[1]) { /* no integer part */ |
| 512 | fout[0] = x[0]; |
| 513 | fout[1] = x[1]; |
| 514 | return NULL; |
| 515 | } |
| 516 | while (!(fh & 0x100000)) { |
| 517 | ex--; |
| 518 | fh = (fh << 1) | ((fl >> 31) & 1); |
| 519 | fl = (fl & 0x7FFFFFFF) << 1; |
| 520 | } |
| 521 | fout[0] = sign | (ex << 20) | (fh & 0xFFFFF); |
| 522 | fout[1] = fl; |
| 523 | return NULL; |
| 524 | } |
| 525 | |
| 526 | char *test_modff(uint32 *x, uint32 *fout, uint32 *iout) { |
| 527 | int ex = (*x >> 23) & 0xFF; /* exponent */ |
| 528 | int sign = *x & 0x80000000; |
| 529 | uint32 f; |
| 530 | |
| 531 | if ((*x & 0x7FFFFFFF) > 0x7F800000) { |
| 532 | /* |
| 533 | * NaN input: return the same in _both_ outputs. |
| 534 | */ |
| 535 | *fout = *iout = *x; |
| 536 | return NULL; |
| 537 | } |
| 538 | |
| 539 | test_rintf(x, iout, 0, 0); |
| 540 | f = *x - *iout; |
| 541 | if (!f) { /* no fraction part */ |
| 542 | *fout = sign; |
| 543 | return NULL; |
| 544 | } |
| 545 | if (!(*iout & 0x7FFFFFFF)) { /* no integer part */ |
| 546 | *fout = *x; |
| 547 | return NULL; |
| 548 | } |
| 549 | while (!(f & 0x800000)) { |
| 550 | ex--; |
| 551 | f = f << 1; |
| 552 | } |
| 553 | *fout = sign | (ex << 23) | (f & 0x7FFFFF); |
| 554 | return NULL; |
| 555 | } |
| 556 | |
| 557 | char *test_copysign(uint32 *x, uint32 *y, uint32 *out) |
| 558 | { |
| 559 | int ysign = y[0] & 0x80000000; |
| 560 | int xhigh = x[0] & 0x7fffffff; |
| 561 | |
| 562 | out[0] = ysign | xhigh; |
| 563 | out[1] = x[1]; |
| 564 | |
| 565 | /* There can be no error */ |
| 566 | return NULL; |
| 567 | } |
| 568 | |
| 569 | char *test_copysignf(uint32 *x, uint32 *y, uint32 *out) |
| 570 | { |
| 571 | int ysign = y[0] & 0x80000000; |
| 572 | int xhigh = x[0] & 0x7fffffff; |
| 573 | |
| 574 | out[0] = ysign | xhigh; |
| 575 | |
| 576 | /* There can be no error */ |
| 577 | return NULL; |
| 578 | } |
| 579 | |
| 580 | char *test_isfinite(uint32 *x, uint32 *out) |
| 581 | { |
| 582 | int xhigh = x[0]; |
| 583 | /* Being finite means that the exponent is not 0x7ff */ |
| 584 | if ((xhigh & 0x7ff00000) == 0x7ff00000) out[0] = 0; |
| 585 | else out[0] = 1; |
| 586 | return NULL; |
| 587 | } |
| 588 | |
| 589 | char *test_isfinitef(uint32 *x, uint32 *out) |
| 590 | { |
| 591 | /* Being finite means that the exponent is not 0xff */ |
| 592 | if ((x[0] & 0x7f800000) == 0x7f800000) out[0] = 0; |
| 593 | else out[0] = 1; |
| 594 | return NULL; |
| 595 | } |
| 596 | |
| 597 | char *test_isinff(uint32 *x, uint32 *out) |
| 598 | { |
| 599 | /* Being infinite means that our bottom 30 bits equate to 0x7f800000 */ |
| 600 | if ((x[0] & 0x7fffffff) == 0x7f800000) out[0] = 1; |
| 601 | else out[0] = 0; |
| 602 | return NULL; |
| 603 | } |
| 604 | |
| 605 | char *test_isinf(uint32 *x, uint32 *out) |
| 606 | { |
| 607 | int xhigh = x[0]; |
| 608 | int xlow = x[1]; |
| 609 | /* Being infinite means that our fraction is zero and exponent is 0x7ff */ |
| 610 | if (((xhigh & 0x7fffffff) == 0x7ff00000) && (xlow == 0)) out[0] = 1; |
| 611 | else out[0] = 0; |
| 612 | return NULL; |
| 613 | } |
| 614 | |
| 615 | char *test_isnanf(uint32 *x, uint32 *out) |
| 616 | { |
| 617 | /* Being NaN means that our exponent is 0xff and non-0 fraction */ |
| 618 | int exponent = x[0] & 0x7f800000; |
| 619 | int fraction = x[0] & 0x007fffff; |
| 620 | if ((exponent == 0x7f800000) && (fraction != 0)) out[0] = 1; |
| 621 | else out[0] = 0; |
| 622 | return NULL; |
| 623 | } |
| 624 | |
| 625 | char *test_isnan(uint32 *x, uint32 *out) |
| 626 | { |
| 627 | /* Being NaN means that our exponent is 0x7ff and non-0 fraction */ |
| 628 | int exponent = x[0] & 0x7ff00000; |
| 629 | int fractionhigh = x[0] & 0x000fffff; |
| 630 | if ((exponent == 0x7ff00000) && ((fractionhigh != 0) || x[1] != 0)) |
| 631 | out[0] = 1; |
| 632 | else out[0] = 0; |
| 633 | return NULL; |
| 634 | } |
| 635 | |
| 636 | char *test_isnormalf(uint32 *x, uint32 *out) |
| 637 | { |
| 638 | /* Being normal means exponent is not 0 and is not 0xff */ |
| 639 | int exponent = x[0] & 0x7f800000; |
| 640 | if (exponent == 0x7f800000) out[0] = 0; |
| 641 | else if (exponent == 0) out[0] = 0; |
| 642 | else out[0] = 1; |
| 643 | return NULL; |
| 644 | } |
| 645 | |
| 646 | char *test_isnormal(uint32 *x, uint32 *out) |
| 647 | { |
| 648 | /* Being normal means exponent is not 0 and is not 0x7ff */ |
| 649 | int exponent = x[0] & 0x7ff00000; |
| 650 | if (exponent == 0x7ff00000) out[0] = 0; |
| 651 | else if (exponent == 0) out[0] = 0; |
| 652 | else out[0] = 1; |
| 653 | return NULL; |
| 654 | } |
| 655 | |
| 656 | char *test_signbitf(uint32 *x, uint32 *out) |
| 657 | { |
| 658 | /* Sign bit is bit 31 */ |
| 659 | out[0] = (x[0] >> 31) & 1; |
| 660 | return NULL; |
| 661 | } |
| 662 | |
| 663 | char *test_signbit(uint32 *x, uint32 *out) |
| 664 | { |
| 665 | /* Sign bit is bit 31 */ |
| 666 | out[0] = (x[0] >> 31) & 1; |
| 667 | return NULL; |
| 668 | } |
| 669 | |
| 670 | char *test_fpclassify(uint32 *x, uint32 *out) |
| 671 | { |
| 672 | int exponent = (x[0] & 0x7ff00000) >> 20; |
| 673 | int fraction = (x[0] & 0x000fffff) | x[1]; |
| 674 | |
| 675 | if ((exponent == 0x00) && (fraction == 0)) out[0] = 0; |
| 676 | else if ((exponent == 0x00) && (fraction != 0)) out[0] = 4; |
| 677 | else if ((exponent == 0x7ff) && (fraction == 0)) out[0] = 3; |
| 678 | else if ((exponent == 0x7ff) && (fraction != 0)) out[0] = 7; |
| 679 | else out[0] = 5; |
| 680 | return NULL; |
| 681 | } |
| 682 | |
| 683 | char *test_fpclassifyf(uint32 *x, uint32 *out) |
| 684 | { |
| 685 | int exponent = (x[0] & 0x7f800000) >> 23; |
| 686 | int fraction = x[0] & 0x007fffff; |
| 687 | |
| 688 | if ((exponent == 0x000) && (fraction == 0)) out[0] = 0; |
| 689 | else if ((exponent == 0x000) && (fraction != 0)) out[0] = 4; |
| 690 | else if ((exponent == 0xff) && (fraction == 0)) out[0] = 3; |
| 691 | else if ((exponent == 0xff) && (fraction != 0)) out[0] = 7; |
| 692 | else out[0] = 5; |
| 693 | return NULL; |
| 694 | } |
| 695 | |
| 696 | /* |
| 697 | * Internal function that compares doubles in x & y and returns -3, -2, -1, 0, |
| 698 | * 1 if they compare to be signaling, unordered, less than, equal or greater |
| 699 | * than. |
| 700 | */ |
| 701 | static int fpcmp4(uint32 *x, uint32 *y) |
| 702 | { |
| 703 | int result = 0; |
| 704 | |
| 705 | /* |
| 706 | * Sort out whether results are ordered or not to begin with |
| 707 | * NaNs have exponent 0x7ff, and non-zero fraction. Signaling NaNs take |
| 708 | * higher priority than quiet ones. |
| 709 | */ |
| 710 | if ((x[0] & 0x7fffffff) >= 0x7ff80000) result = -2; |
| 711 | else if ((x[0] & 0x7fffffff) > 0x7ff00000) result = -3; |
| 712 | else if (((x[0] & 0x7fffffff) == 0x7ff00000) && (x[1] != 0)) result = -3; |
| 713 | if ((y[0] & 0x7fffffff) >= 0x7ff80000 && result != -3) result = -2; |
| 714 | else if ((y[0] & 0x7fffffff) > 0x7ff00000) result = -3; |
| 715 | else if (((y[0] & 0x7fffffff) == 0x7ff00000) && (y[1] != 0)) result = -3; |
| 716 | if (result != 0) return result; |
| 717 | |
| 718 | /* |
| 719 | * The two forms of zero are equal |
| 720 | */ |
| 721 | if (((x[0] & 0x7fffffff) == 0) && x[1] == 0 && |
| 722 | ((y[0] & 0x7fffffff) == 0) && y[1] == 0) |
| 723 | return 0; |
| 724 | |
| 725 | /* |
| 726 | * If x and y have different signs we can tell that they're not equal |
| 727 | * If x is +ve we have x > y return 1 - otherwise y is +ve return -1 |
| 728 | */ |
| 729 | if ((x[0] >> 31) != (y[0] >> 31)) |
| 730 | return ((x[0] >> 31) == 0) - ((y[0] >> 31) == 0); |
| 731 | |
| 732 | /* |
| 733 | * Now we have both signs the same, let's do an initial compare of the |
| 734 | * values. |
| 735 | * |
| 736 | * Whoever designed IEEE754's floating point formats is very clever and |
| 737 | * earns my undying admiration. Once you remove the sign-bit, the |
| 738 | * floating point numbers can be ordered using the standard <, ==, > |
| 739 | * operators will treating the fp-numbers as integers with that bit- |
| 740 | * pattern. |
| 741 | */ |
| 742 | if ((x[0] & 0x7fffffff) < (y[0] & 0x7fffffff)) result = -1; |
| 743 | else if ((x[0] & 0x7fffffff) > (y[0] & 0x7fffffff)) result = 1; |
| 744 | else if (x[1] < y[1]) result = -1; |
| 745 | else if (x[1] > y[1]) result = 1; |
| 746 | else result = 0; |
| 747 | |
| 748 | /* |
| 749 | * Now we return the result - is x is positive (and therefore so is y) we |
| 750 | * return the plain result - otherwise we negate it and return. |
| 751 | */ |
| 752 | if ((x[0] >> 31) == 0) return result; |
| 753 | else return -result; |
| 754 | } |
| 755 | |
| 756 | /* |
| 757 | * Internal function that compares floats in x & y and returns -3, -2, -1, 0, |
| 758 | * 1 if they compare to be signaling, unordered, less than, equal or greater |
| 759 | * than. |
| 760 | */ |
| 761 | static int fpcmp4f(uint32 *x, uint32 *y) |
| 762 | { |
| 763 | int result = 0; |
| 764 | |
| 765 | /* |
| 766 | * Sort out whether results are ordered or not to begin with |
| 767 | * NaNs have exponent 0xff, and non-zero fraction - we have to handle all |
| 768 | * signaling cases over the quiet ones |
| 769 | */ |
| 770 | if ((x[0] & 0x7fffffff) >= 0x7fc00000) result = -2; |
| 771 | else if ((x[0] & 0x7fffffff) > 0x7f800000) result = -3; |
| 772 | if ((y[0] & 0x7fffffff) >= 0x7fc00000 && result != -3) result = -2; |
| 773 | else if ((y[0] & 0x7fffffff) > 0x7f800000) result = -3; |
| 774 | if (result != 0) return result; |
| 775 | |
| 776 | /* |
| 777 | * The two forms of zero are equal |
| 778 | */ |
| 779 | if (((x[0] & 0x7fffffff) == 0) && ((y[0] & 0x7fffffff) == 0)) |
| 780 | return 0; |
| 781 | |
| 782 | /* |
| 783 | * If x and y have different signs we can tell that they're not equal |
| 784 | * If x is +ve we have x > y return 1 - otherwise y is +ve return -1 |
| 785 | */ |
| 786 | if ((x[0] >> 31) != (y[0] >> 31)) |
| 787 | return ((x[0] >> 31) == 0) - ((y[0] >> 31) == 0); |
| 788 | |
| 789 | /* |
| 790 | * Now we have both signs the same, let's do an initial compare of the |
| 791 | * values. |
| 792 | * |
| 793 | * Whoever designed IEEE754's floating point formats is very clever and |
| 794 | * earns my undying admiration. Once you remove the sign-bit, the |
| 795 | * floating point numbers can be ordered using the standard <, ==, > |
| 796 | * operators will treating the fp-numbers as integers with that bit- |
| 797 | * pattern. |
| 798 | */ |
| 799 | if ((x[0] & 0x7fffffff) < (y[0] & 0x7fffffff)) result = -1; |
| 800 | else if ((x[0] & 0x7fffffff) > (y[0] & 0x7fffffff)) result = 1; |
| 801 | else result = 0; |
| 802 | |
| 803 | /* |
| 804 | * Now we return the result - is x is positive (and therefore so is y) we |
| 805 | * return the plain result - otherwise we negate it and return. |
| 806 | */ |
| 807 | if ((x[0] >> 31) == 0) return result; |
| 808 | else return -result; |
| 809 | } |
| 810 | |
| 811 | char *test_isgreater(uint32 *x, uint32 *y, uint32 *out) |
| 812 | { |
| 813 | int result = fpcmp4(x, y); |
| 814 | *out = (result == 1); |
| 815 | return result == -3 ? "i" : NULL; |
| 816 | } |
| 817 | |
| 818 | char *test_isgreaterequal(uint32 *x, uint32 *y, uint32 *out) |
| 819 | { |
| 820 | int result = fpcmp4(x, y); |
| 821 | *out = (result >= 0); |
| 822 | return result == -3 ? "i" : NULL; |
| 823 | } |
| 824 | |
| 825 | char *test_isless(uint32 *x, uint32 *y, uint32 *out) |
| 826 | { |
| 827 | int result = fpcmp4(x, y); |
| 828 | *out = (result == -1); |
| 829 | return result == -3 ? "i" : NULL; |
| 830 | } |
| 831 | |
| 832 | char *test_islessequal(uint32 *x, uint32 *y, uint32 *out) |
| 833 | { |
| 834 | int result = fpcmp4(x, y); |
| 835 | *out = (result == -1) || (result == 0); |
| 836 | return result == -3 ? "i" : NULL; |
| 837 | } |
| 838 | |
| 839 | char *test_islessgreater(uint32 *x, uint32 *y, uint32 *out) |
| 840 | { |
| 841 | int result = fpcmp4(x, y); |
| 842 | *out = (result == -1) || (result == 1); |
| 843 | return result == -3 ? "i" : NULL; |
| 844 | } |
| 845 | |
| 846 | char *test_isunordered(uint32 *x, uint32 *y, uint32 *out) |
| 847 | { |
| 848 | int normal = 0; |
| 849 | int result = fpcmp4(x, y); |
| 850 | |
| 851 | test_isnormal(x, out); |
| 852 | normal |= *out; |
| 853 | test_isnormal(y, out); |
| 854 | normal |= *out; |
| 855 | *out = (result == -2) || (result == -3); |
| 856 | return result == -3 ? "i" : NULL; |
| 857 | } |
| 858 | |
| 859 | char *test_isgreaterf(uint32 *x, uint32 *y, uint32 *out) |
| 860 | { |
| 861 | int result = fpcmp4f(x, y); |
| 862 | *out = (result == 1); |
| 863 | return result == -3 ? "i" : NULL; |
| 864 | } |
| 865 | |
| 866 | char *test_isgreaterequalf(uint32 *x, uint32 *y, uint32 *out) |
| 867 | { |
| 868 | int result = fpcmp4f(x, y); |
| 869 | *out = (result >= 0); |
| 870 | return result == -3 ? "i" : NULL; |
| 871 | } |
| 872 | |
| 873 | char *test_islessf(uint32 *x, uint32 *y, uint32 *out) |
| 874 | { |
| 875 | int result = fpcmp4f(x, y); |
| 876 | *out = (result == -1); |
| 877 | return result == -3 ? "i" : NULL; |
| 878 | } |
| 879 | |
| 880 | char *test_islessequalf(uint32 *x, uint32 *y, uint32 *out) |
| 881 | { |
| 882 | int result = fpcmp4f(x, y); |
| 883 | *out = (result == -1) || (result == 0); |
| 884 | return result == -3 ? "i" : NULL; |
| 885 | } |
| 886 | |
| 887 | char *test_islessgreaterf(uint32 *x, uint32 *y, uint32 *out) |
| 888 | { |
| 889 | int result = fpcmp4f(x, y); |
| 890 | *out = (result == -1) || (result == 1); |
| 891 | return result == -3 ? "i" : NULL; |
| 892 | } |
| 893 | |
| 894 | char *test_isunorderedf(uint32 *x, uint32 *y, uint32 *out) |
| 895 | { |
| 896 | int normal = 0; |
| 897 | int result = fpcmp4f(x, y); |
| 898 | |
| 899 | test_isnormalf(x, out); |
| 900 | normal |= *out; |
| 901 | test_isnormalf(y, out); |
| 902 | normal |= *out; |
| 903 | *out = (result == -2) || (result == -3); |
| 904 | return result == -3 ? "i" : NULL; |
| 905 | } |