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 | * dotest.c - actually generate mathlib test cases |
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 <string.h> |
| 10 | #include <stdlib.h> |
| 11 | #include <stdint.h> |
| 12 | #include <assert.h> |
| 13 | #include <limits.h> |
| 14 | |
| 15 | #include "semi.h" |
| 16 | #include "intern.h" |
| 17 | #include "random.h" |
| 18 | |
| 19 | #define MPFR_PREC 96 /* good enough for float or double + a few extra bits */ |
| 20 | |
| 21 | extern int lib_fo, lib_no_arith, ntests; |
| 22 | |
| 23 | /* |
| 24 | * Prototypes. |
| 25 | */ |
| 26 | static void cases_biased(uint32 *, uint32, uint32); |
| 27 | static void cases_biased_positive(uint32 *, uint32, uint32); |
| 28 | static void cases_biased_float(uint32 *, uint32, uint32); |
| 29 | static void cases_uniform(uint32 *, uint32, uint32); |
| 30 | static void cases_uniform_positive(uint32 *, uint32, uint32); |
| 31 | static void cases_uniform_float(uint32 *, uint32, uint32); |
| 32 | static void cases_uniform_float_positive(uint32 *, uint32, uint32); |
| 33 | static void log_cases(uint32 *, uint32, uint32); |
| 34 | static void log_cases_float(uint32 *, uint32, uint32); |
| 35 | static void log1p_cases(uint32 *, uint32, uint32); |
| 36 | static void log1p_cases_float(uint32 *, uint32, uint32); |
| 37 | static void minmax_cases(uint32 *, uint32, uint32); |
| 38 | static void minmax_cases_float(uint32 *, uint32, uint32); |
| 39 | static void atan2_cases(uint32 *, uint32, uint32); |
| 40 | static void atan2_cases_float(uint32 *, uint32, uint32); |
| 41 | static void pow_cases(uint32 *, uint32, uint32); |
| 42 | static void pow_cases_float(uint32 *, uint32, uint32); |
| 43 | static void rred_cases(uint32 *, uint32, uint32); |
| 44 | static void rred_cases_float(uint32 *, uint32, uint32); |
| 45 | static void cases_semi1(uint32 *, uint32, uint32); |
| 46 | static void cases_semi1_float(uint32 *, uint32, uint32); |
| 47 | static void cases_semi2(uint32 *, uint32, uint32); |
| 48 | static void cases_semi2_float(uint32 *, uint32, uint32); |
| 49 | static void cases_ldexp(uint32 *, uint32, uint32); |
| 50 | static void cases_ldexp_float(uint32 *, uint32, uint32); |
| 51 | |
| 52 | static void complex_cases_uniform(uint32 *, uint32, uint32); |
| 53 | static void complex_cases_uniform_float(uint32 *, uint32, uint32); |
| 54 | static void complex_cases_biased(uint32 *, uint32, uint32); |
| 55 | static void complex_cases_biased_float(uint32 *, uint32, uint32); |
| 56 | static void complex_log_cases(uint32 *, uint32, uint32); |
| 57 | static void complex_log_cases_float(uint32 *, uint32, uint32); |
| 58 | static void complex_pow_cases(uint32 *, uint32, uint32); |
| 59 | static void complex_pow_cases_float(uint32 *, uint32, uint32); |
| 60 | static void complex_arithmetic_cases(uint32 *, uint32, uint32); |
| 61 | static void complex_arithmetic_cases_float(uint32 *, uint32, uint32); |
| 62 | |
| 63 | static uint32 doubletop(int x, int scale); |
| 64 | static uint32 floatval(int x, int scale); |
| 65 | |
| 66 | /* |
| 67 | * Convert back and forth between IEEE bit patterns and the |
| 68 | * mpfr_t/mpc_t types. |
| 69 | */ |
| 70 | static void set_mpfr_d(mpfr_t x, uint32 h, uint32 l) |
| 71 | { |
| 72 | uint64_t hl = ((uint64_t)h << 32) | l; |
| 73 | uint32 exp = (hl >> 52) & 0x7ff; |
| 74 | int64_t mantissa = hl & (((uint64_t)1 << 52) - 1); |
| 75 | int sign = (hl >> 63) ? -1 : +1; |
| 76 | if (exp == 0x7ff) { |
| 77 | if (mantissa == 0) |
| 78 | mpfr_set_inf(x, sign); |
| 79 | else |
| 80 | mpfr_set_nan(x); |
| 81 | } else if (exp == 0 && mantissa == 0) { |
| 82 | mpfr_set_ui(x, 0, GMP_RNDN); |
| 83 | mpfr_setsign(x, x, sign < 0, GMP_RNDN); |
| 84 | } else { |
| 85 | if (exp != 0) |
| 86 | mantissa |= ((uint64_t)1 << 52); |
| 87 | else |
| 88 | exp++; |
| 89 | mpfr_set_sj_2exp(x, mantissa * sign, (int)exp - 0x3ff - 52, GMP_RNDN); |
| 90 | } |
| 91 | } |
| 92 | static void set_mpfr_f(mpfr_t x, uint32 f) |
| 93 | { |
| 94 | uint32 exp = (f >> 23) & 0xff; |
| 95 | int32 mantissa = f & ((1 << 23) - 1); |
| 96 | int sign = (f >> 31) ? -1 : +1; |
| 97 | if (exp == 0xff) { |
| 98 | if (mantissa == 0) |
| 99 | mpfr_set_inf(x, sign); |
| 100 | else |
| 101 | mpfr_set_nan(x); |
| 102 | } else if (exp == 0 && mantissa == 0) { |
| 103 | mpfr_set_ui(x, 0, GMP_RNDN); |
| 104 | mpfr_setsign(x, x, sign < 0, GMP_RNDN); |
| 105 | } else { |
| 106 | if (exp != 0) |
| 107 | mantissa |= (1 << 23); |
| 108 | else |
| 109 | exp++; |
| 110 | mpfr_set_sj_2exp(x, mantissa * sign, (int)exp - 0x7f - 23, GMP_RNDN); |
| 111 | } |
| 112 | } |
| 113 | static void set_mpc_d(mpc_t z, uint32 rh, uint32 rl, uint32 ih, uint32 il) |
| 114 | { |
| 115 | mpfr_t x, y; |
| 116 | mpfr_init2(x, MPFR_PREC); |
| 117 | mpfr_init2(y, MPFR_PREC); |
| 118 | set_mpfr_d(x, rh, rl); |
| 119 | set_mpfr_d(y, ih, il); |
| 120 | mpc_set_fr_fr(z, x, y, MPC_RNDNN); |
| 121 | mpfr_clear(x); |
| 122 | mpfr_clear(y); |
| 123 | } |
| 124 | static void set_mpc_f(mpc_t z, uint32 r, uint32 i) |
| 125 | { |
| 126 | mpfr_t x, y; |
| 127 | mpfr_init2(x, MPFR_PREC); |
| 128 | mpfr_init2(y, MPFR_PREC); |
| 129 | set_mpfr_f(x, r); |
| 130 | set_mpfr_f(y, i); |
| 131 | mpc_set_fr_fr(z, x, y, MPC_RNDNN); |
| 132 | mpfr_clear(x); |
| 133 | mpfr_clear(y); |
| 134 | } |
| 135 | static void get_mpfr_d(const mpfr_t x, uint32 *h, uint32 *l, uint32 *extra) |
| 136 | { |
| 137 | uint32_t sign, expfield, mantfield; |
| 138 | mpfr_t significand; |
| 139 | int exp; |
| 140 | |
| 141 | if (mpfr_nan_p(x)) { |
| 142 | *h = 0x7ff80000; |
| 143 | *l = 0; |
| 144 | *extra = 0; |
| 145 | return; |
| 146 | } |
| 147 | |
| 148 | sign = mpfr_signbit(x) ? 0x80000000U : 0; |
| 149 | |
| 150 | if (mpfr_inf_p(x)) { |
| 151 | *h = 0x7ff00000 | sign; |
| 152 | *l = 0; |
| 153 | *extra = 0; |
| 154 | return; |
| 155 | } |
| 156 | |
| 157 | if (mpfr_zero_p(x)) { |
| 158 | *h = 0x00000000 | sign; |
| 159 | *l = 0; |
| 160 | *extra = 0; |
| 161 | return; |
| 162 | } |
| 163 | |
| 164 | mpfr_init2(significand, MPFR_PREC); |
| 165 | mpfr_set(significand, x, GMP_RNDN); |
| 166 | exp = mpfr_get_exp(significand); |
| 167 | mpfr_set_exp(significand, 0); |
| 168 | |
| 169 | /* Now significand is in [1/2,1), and significand * 2^exp == x. |
| 170 | * So the IEEE exponent corresponding to exp==0 is 0x3fe. */ |
| 171 | if (exp > 0x400) { |
| 172 | /* overflow to infinity anyway */ |
| 173 | *h = 0x7ff00000 | sign; |
| 174 | *l = 0; |
| 175 | *extra = 0; |
| 176 | mpfr_clear(significand); |
| 177 | return; |
| 178 | } |
| 179 | |
| 180 | if (exp <= -0x3fe || mpfr_zero_p(x)) |
| 181 | exp = -0x3fd; /* denormalise */ |
| 182 | expfield = exp + 0x3fd; /* offset to cancel leading mantissa bit */ |
| 183 | |
| 184 | mpfr_div_2si(significand, x, exp - 21, GMP_RNDN); |
| 185 | mpfr_abs(significand, significand, GMP_RNDN); |
| 186 | mantfield = mpfr_get_ui(significand, GMP_RNDZ); |
| 187 | *h = sign + ((uint64_t)expfield << 20) + mantfield; |
| 188 | mpfr_sub_ui(significand, significand, mantfield, GMP_RNDN); |
| 189 | mpfr_mul_2ui(significand, significand, 32, GMP_RNDN); |
| 190 | mantfield = mpfr_get_ui(significand, GMP_RNDZ); |
| 191 | *l = mantfield; |
| 192 | mpfr_sub_ui(significand, significand, mantfield, GMP_RNDN); |
| 193 | mpfr_mul_2ui(significand, significand, 32, GMP_RNDN); |
| 194 | mantfield = mpfr_get_ui(significand, GMP_RNDZ); |
| 195 | *extra = mantfield; |
| 196 | |
| 197 | mpfr_clear(significand); |
| 198 | } |
| 199 | static void get_mpfr_f(const mpfr_t x, uint32 *f, uint32 *extra) |
| 200 | { |
| 201 | uint32_t sign, expfield, mantfield; |
| 202 | mpfr_t significand; |
| 203 | int exp; |
| 204 | |
| 205 | if (mpfr_nan_p(x)) { |
| 206 | *f = 0x7fc00000; |
| 207 | *extra = 0; |
| 208 | return; |
| 209 | } |
| 210 | |
| 211 | sign = mpfr_signbit(x) ? 0x80000000U : 0; |
| 212 | |
| 213 | if (mpfr_inf_p(x)) { |
| 214 | *f = 0x7f800000 | sign; |
| 215 | *extra = 0; |
| 216 | return; |
| 217 | } |
| 218 | |
| 219 | if (mpfr_zero_p(x)) { |
| 220 | *f = 0x00000000 | sign; |
| 221 | *extra = 0; |
| 222 | return; |
| 223 | } |
| 224 | |
| 225 | mpfr_init2(significand, MPFR_PREC); |
| 226 | mpfr_set(significand, x, GMP_RNDN); |
| 227 | exp = mpfr_get_exp(significand); |
| 228 | mpfr_set_exp(significand, 0); |
| 229 | |
| 230 | /* Now significand is in [1/2,1), and significand * 2^exp == x. |
| 231 | * So the IEEE exponent corresponding to exp==0 is 0x7e. */ |
| 232 | if (exp > 0x80) { |
| 233 | /* overflow to infinity anyway */ |
| 234 | *f = 0x7f800000 | sign; |
| 235 | *extra = 0; |
| 236 | mpfr_clear(significand); |
| 237 | return; |
| 238 | } |
| 239 | |
| 240 | if (exp <= -0x7e || mpfr_zero_p(x)) |
| 241 | exp = -0x7d; /* denormalise */ |
| 242 | expfield = exp + 0x7d; /* offset to cancel leading mantissa bit */ |
| 243 | |
| 244 | mpfr_div_2si(significand, x, exp - 24, GMP_RNDN); |
| 245 | mpfr_abs(significand, significand, GMP_RNDN); |
| 246 | mantfield = mpfr_get_ui(significand, GMP_RNDZ); |
| 247 | *f = sign + ((uint64_t)expfield << 23) + mantfield; |
| 248 | mpfr_sub_ui(significand, significand, mantfield, GMP_RNDN); |
| 249 | mpfr_mul_2ui(significand, significand, 32, GMP_RNDN); |
| 250 | mantfield = mpfr_get_ui(significand, GMP_RNDZ); |
| 251 | *extra = mantfield; |
| 252 | |
| 253 | mpfr_clear(significand); |
| 254 | } |
| 255 | static void get_mpc_d(const mpc_t z, |
| 256 | uint32 *rh, uint32 *rl, uint32 *rextra, |
| 257 | uint32 *ih, uint32 *il, uint32 *iextra) |
| 258 | { |
| 259 | mpfr_t x, y; |
| 260 | mpfr_init2(x, MPFR_PREC); |
| 261 | mpfr_init2(y, MPFR_PREC); |
| 262 | mpc_real(x, z, GMP_RNDN); |
| 263 | mpc_imag(y, z, GMP_RNDN); |
| 264 | get_mpfr_d(x, rh, rl, rextra); |
| 265 | get_mpfr_d(y, ih, il, iextra); |
| 266 | mpfr_clear(x); |
| 267 | mpfr_clear(y); |
| 268 | } |
| 269 | static void get_mpc_f(const mpc_t z, |
| 270 | uint32 *r, uint32 *rextra, |
| 271 | uint32 *i, uint32 *iextra) |
| 272 | { |
| 273 | mpfr_t x, y; |
| 274 | mpfr_init2(x, MPFR_PREC); |
| 275 | mpfr_init2(y, MPFR_PREC); |
| 276 | mpc_real(x, z, GMP_RNDN); |
| 277 | mpc_imag(y, z, GMP_RNDN); |
| 278 | get_mpfr_f(x, r, rextra); |
| 279 | get_mpfr_f(y, i, iextra); |
| 280 | mpfr_clear(x); |
| 281 | mpfr_clear(y); |
| 282 | } |
| 283 | |
| 284 | /* |
| 285 | * Implementation of mathlib functions that aren't trivially |
| 286 | * implementable using an existing mpfr or mpc function. |
| 287 | */ |
| 288 | int test_rred(mpfr_t ret, const mpfr_t x, int *quadrant) |
| 289 | { |
| 290 | mpfr_t halfpi; |
| 291 | long quo; |
| 292 | int status; |
| 293 | |
| 294 | /* |
| 295 | * In the worst case of range reduction, we get an input of size |
| 296 | * around 2^1024, and must find its remainder mod pi, which means |
| 297 | * we need 1024 bits of pi at least. Plus, the remainder might |
| 298 | * happen to come out very very small if we're unlucky. How |
| 299 | * unlucky can we be? Well, conveniently, I once went through and |
| 300 | * actually worked that out using Paxson's modular minimisation |
| 301 | * algorithm, and it turns out that the smallest exponent you can |
| 302 | * get out of a nontrivial[1] double precision range reduction is |
| 303 | * 0x3c2, i.e. of the order of 2^-61. So we need 1024 bits of pi |
| 304 | * to get us down to the units digit, another 61 or so bits (say |
| 305 | * 64) to get down to the highest set bit of the output, and then |
| 306 | * some bits to make the actual mantissa big enough. |
| 307 | * |
| 308 | * [1] of course the output of range reduction can have an |
| 309 | * arbitrarily small exponent in the trivial case, where the |
| 310 | * input is so small that it's the identity function. That |
| 311 | * doesn't count. |
| 312 | */ |
| 313 | mpfr_init2(halfpi, MPFR_PREC + 1024 + 64); |
| 314 | mpfr_const_pi(halfpi, GMP_RNDN); |
| 315 | mpfr_div_ui(halfpi, halfpi, 2, GMP_RNDN); |
| 316 | |
| 317 | status = mpfr_remquo(ret, &quo, x, halfpi, GMP_RNDN); |
| 318 | *quadrant = quo & 3; |
| 319 | |
| 320 | mpfr_clear(halfpi); |
| 321 | |
| 322 | return status; |
| 323 | } |
| 324 | int test_lgamma(mpfr_t ret, const mpfr_t x, mpfr_rnd_t rnd) |
| 325 | { |
| 326 | /* |
| 327 | * mpfr_lgamma takes an extra int * parameter to hold the output |
| 328 | * sign. We don't bother testing that, so this wrapper throws away |
| 329 | * the sign and hence fits into the same function prototype as all |
| 330 | * the other real->real mpfr functions. |
| 331 | * |
| 332 | * There is also mpfr_lngamma which has no sign output and hence |
| 333 | * has the right prototype already, but unfortunately it returns |
| 334 | * NaN in cases where gamma(x) < 0, so it's no use to us. |
| 335 | */ |
| 336 | int sign; |
| 337 | return mpfr_lgamma(ret, &sign, x, rnd); |
| 338 | } |
| 339 | int test_cpow(mpc_t ret, const mpc_t x, const mpc_t y, mpc_rnd_t rnd) |
| 340 | { |
| 341 | /* |
| 342 | * For complex pow, we must bump up the precision by a huge amount |
| 343 | * if we want it to get the really difficult cases right. (Not |
| 344 | * that we expect the library under test to be getting those cases |
| 345 | * right itself, but we'd at least like the test suite to report |
| 346 | * them as wrong for the _right reason_.) |
| 347 | * |
| 348 | * This works around a bug in mpc_pow(), fixed by r1455 in the MPC |
| 349 | * svn repository (2014-10-14) and expected to be in any MPC |
| 350 | * release after 1.0.2 (which was the latest release already made |
| 351 | * at the time of the fix). So as and when we update to an MPC |
| 352 | * with the fix in it, we could remove this workaround. |
| 353 | * |
| 354 | * For the reasons for choosing this amount of extra precision, |
| 355 | * see analysis in complex/cpownotes.txt for the rationale for the |
| 356 | * amount. |
| 357 | */ |
| 358 | mpc_t xbig, ybig, retbig; |
| 359 | int status; |
| 360 | |
| 361 | mpc_init2(xbig, 1034 + 53 + 60 + MPFR_PREC); |
| 362 | mpc_init2(ybig, 1034 + 53 + 60 + MPFR_PREC); |
| 363 | mpc_init2(retbig, 1034 + 53 + 60 + MPFR_PREC); |
| 364 | |
| 365 | mpc_set(xbig, x, MPC_RNDNN); |
| 366 | mpc_set(ybig, y, MPC_RNDNN); |
| 367 | status = mpc_pow(retbig, xbig, ybig, rnd); |
| 368 | mpc_set(ret, retbig, rnd); |
| 369 | |
| 370 | mpc_clear(xbig); |
| 371 | mpc_clear(ybig); |
| 372 | mpc_clear(retbig); |
| 373 | |
| 374 | return status; |
| 375 | } |
| 376 | |
| 377 | /* |
| 378 | * Identify 'hard' values (NaN, Inf, nonzero denormal) for deciding |
| 379 | * whether microlib will decline to run a test. |
| 380 | */ |
| 381 | #define is_shard(in) ( \ |
| 382 | (((in)[0] & 0x7F800000) == 0x7F800000 || \ |
| 383 | (((in)[0] & 0x7F800000) == 0 && ((in)[0]&0x7FFFFFFF) != 0))) |
| 384 | |
| 385 | #define is_dhard(in) ( \ |
| 386 | (((in)[0] & 0x7FF00000) == 0x7FF00000 || \ |
| 387 | (((in)[0] & 0x7FF00000) == 0 && (((in)[0] & 0xFFFFF) | (in)[1]) != 0))) |
| 388 | |
| 389 | /* |
| 390 | * Identify integers. |
| 391 | */ |
| 392 | int is_dinteger(uint32 *in) |
| 393 | { |
| 394 | uint32 out[3]; |
| 395 | if ((0x7FF00000 & ~in[0]) == 0) |
| 396 | return 0; /* not finite, hence not integer */ |
| 397 | test_ceil(in, out); |
| 398 | return in[0] == out[0] && in[1] == out[1]; |
| 399 | } |
| 400 | int is_sinteger(uint32 *in) |
| 401 | { |
| 402 | uint32 out[3]; |
| 403 | if ((0x7F800000 & ~in[0]) == 0) |
| 404 | return 0; /* not finite, hence not integer */ |
| 405 | test_ceilf(in, out); |
| 406 | return in[0] == out[0]; |
| 407 | } |
| 408 | |
| 409 | /* |
| 410 | * Identify signalling NaNs. |
| 411 | */ |
| 412 | int is_dsnan(const uint32 *in) |
| 413 | { |
| 414 | if ((in[0] & 0x7FF00000) != 0x7FF00000) |
| 415 | return 0; /* not the inf/nan exponent */ |
| 416 | if ((in[0] << 12) == 0 && in[1] == 0) |
| 417 | return 0; /* inf */ |
| 418 | if (in[0] & 0x00080000) |
| 419 | return 0; /* qnan */ |
| 420 | return 1; |
| 421 | } |
| 422 | int is_ssnan(const uint32 *in) |
| 423 | { |
| 424 | if ((in[0] & 0x7F800000) != 0x7F800000) |
| 425 | return 0; /* not the inf/nan exponent */ |
| 426 | if ((in[0] << 9) == 0) |
| 427 | return 0; /* inf */ |
| 428 | if (in[0] & 0x00400000) |
| 429 | return 0; /* qnan */ |
| 430 | return 1; |
| 431 | } |
| 432 | int is_snan(const uint32 *in, int size) |
| 433 | { |
| 434 | return size == 2 ? is_dsnan(in) : is_ssnan(in); |
| 435 | } |
| 436 | |
| 437 | /* |
| 438 | * Wrapper functions called to fix up unusual results after the main |
| 439 | * test function has run. |
| 440 | */ |
| 441 | void universal_wrapper(wrapperctx *ctx) |
| 442 | { |
| 443 | /* |
| 444 | * Any SNaN input gives rise to a QNaN output. |
| 445 | */ |
| 446 | int op; |
| 447 | for (op = 0; op < wrapper_get_nops(ctx); op++) { |
| 448 | int size = wrapper_get_size(ctx, op); |
| 449 | |
| 450 | if (!wrapper_is_complex(ctx, op) && |
| 451 | is_snan(wrapper_get_ieee(ctx, op), size)) { |
| 452 | wrapper_set_nan(ctx); |
| 453 | } |
| 454 | } |
| 455 | } |
| 456 | |
| 457 | Testable functions[] = { |
| 458 | /* |
| 459 | * Trig functions: sin, cos, tan. We test the core function |
| 460 | * between -16 and +16: we assume that range reduction exists |
| 461 | * and will be used for larger arguments, and we'll test that |
| 462 | * separately. Also we only go down to 2^-27 in magnitude, |
| 463 | * because below that sin(x)=tan(x)=x and cos(x)=1 as far as |
| 464 | * double precision can tell, which is boring. |
| 465 | */ |
| 466 | {"sin", (funcptr)mpfr_sin, args1, {NULL}, |
| 467 | cases_uniform, 0x3e400000, 0x40300000}, |
| 468 | {"sinf", (funcptr)mpfr_sin, args1f, {NULL}, |
| 469 | cases_uniform_float, 0x39800000, 0x41800000}, |
| 470 | {"cos", (funcptr)mpfr_cos, args1, {NULL}, |
| 471 | cases_uniform, 0x3e400000, 0x40300000}, |
| 472 | {"cosf", (funcptr)mpfr_cos, args1f, {NULL}, |
| 473 | cases_uniform_float, 0x39800000, 0x41800000}, |
| 474 | {"tan", (funcptr)mpfr_tan, args1, {NULL}, |
| 475 | cases_uniform, 0x3e400000, 0x40300000}, |
| 476 | {"tanf", (funcptr)mpfr_tan, args1f, {NULL}, |
| 477 | cases_uniform_float, 0x39800000, 0x41800000}, |
Szabolcs Nagy | 5dd369f | 2018-05-15 15:13:42 +0100 | [diff] [blame] | 478 | {"sincosf_sinf", (funcptr)mpfr_sin, args1f, {NULL}, |
| 479 | cases_uniform_float, 0x39800000, 0x41800000}, |
| 480 | {"sincosf_cosf", (funcptr)mpfr_cos, args1f, {NULL}, |
| 481 | cases_uniform_float, 0x39800000, 0x41800000}, |
George Lander | da55ef9 | 2015-11-19 12:05:06 +0000 | [diff] [blame] | 482 | /* |
| 483 | * Inverse trig: asin, acos. Between 1 and -1, of course. acos |
| 484 | * goes down to 2^-54, asin to 2^-27. |
| 485 | */ |
| 486 | {"asin", (funcptr)mpfr_asin, args1, {NULL}, |
| 487 | cases_uniform, 0x3e400000, 0x3fefffff}, |
| 488 | {"asinf", (funcptr)mpfr_asin, args1f, {NULL}, |
| 489 | cases_uniform_float, 0x39800000, 0x3f7fffff}, |
| 490 | {"acos", (funcptr)mpfr_acos, args1, {NULL}, |
| 491 | cases_uniform, 0x3c900000, 0x3fefffff}, |
| 492 | {"acosf", (funcptr)mpfr_acos, args1f, {NULL}, |
| 493 | cases_uniform_float, 0x33800000, 0x3f7fffff}, |
| 494 | /* |
| 495 | * Inverse trig: atan. atan is stable (in double prec) with |
| 496 | * argument magnitude past 2^53, so we'll test up to there. |
| 497 | * atan(x) is boringly just x below 2^-27. |
| 498 | */ |
| 499 | {"atan", (funcptr)mpfr_atan, args1, {NULL}, |
| 500 | cases_uniform, 0x3e400000, 0x43400000}, |
| 501 | {"atanf", (funcptr)mpfr_atan, args1f, {NULL}, |
| 502 | cases_uniform_float, 0x39800000, 0x4b800000}, |
| 503 | /* |
| 504 | * atan2. Interesting cases arise when the exponents of the |
| 505 | * arguments differ by at most about 50. |
| 506 | */ |
| 507 | {"atan2", (funcptr)mpfr_atan2, args2, {NULL}, |
| 508 | atan2_cases, 0}, |
| 509 | {"atan2f", (funcptr)mpfr_atan2, args2f, {NULL}, |
| 510 | atan2_cases_float, 0}, |
| 511 | /* |
| 512 | * The exponentials: exp, sinh, cosh. They overflow at around |
| 513 | * 710. exp and sinh are boring below 2^-54, cosh below 2^-27. |
| 514 | */ |
| 515 | {"exp", (funcptr)mpfr_exp, args1, {NULL}, |
| 516 | cases_uniform, 0x3c900000, 0x40878000}, |
| 517 | {"expf", (funcptr)mpfr_exp, args1f, {NULL}, |
| 518 | cases_uniform_float, 0x33800000, 0x42dc0000}, |
| 519 | {"sinh", (funcptr)mpfr_sinh, args1, {NULL}, |
| 520 | cases_uniform, 0x3c900000, 0x40878000}, |
| 521 | {"sinhf", (funcptr)mpfr_sinh, args1f, {NULL}, |
| 522 | cases_uniform_float, 0x33800000, 0x42dc0000}, |
| 523 | {"cosh", (funcptr)mpfr_cosh, args1, {NULL}, |
| 524 | cases_uniform, 0x3e400000, 0x40878000}, |
| 525 | {"coshf", (funcptr)mpfr_cosh, args1f, {NULL}, |
| 526 | cases_uniform_float, 0x39800000, 0x42dc0000}, |
| 527 | /* |
| 528 | * tanh is stable past around 20. It's boring below 2^-27. |
| 529 | */ |
| 530 | {"tanh", (funcptr)mpfr_tanh, args1, {NULL}, |
| 531 | cases_uniform, 0x3e400000, 0x40340000}, |
| 532 | {"tanhf", (funcptr)mpfr_tanh, args1f, {NULL}, |
| 533 | cases_uniform, 0x39800000, 0x41100000}, |
| 534 | /* |
| 535 | * log must be tested only on positive numbers, but can cover |
| 536 | * the whole range of positive nonzero finite numbers. It never |
| 537 | * gets boring. |
| 538 | */ |
| 539 | {"log", (funcptr)mpfr_log, args1, {NULL}, log_cases, 0}, |
| 540 | {"logf", (funcptr)mpfr_log, args1f, {NULL}, log_cases_float, 0}, |
| 541 | {"log10", (funcptr)mpfr_log10, args1, {NULL}, log_cases, 0}, |
| 542 | {"log10f", (funcptr)mpfr_log10, args1f, {NULL}, log_cases_float, 0}, |
| 543 | /* |
| 544 | * pow. |
| 545 | */ |
| 546 | {"pow", (funcptr)mpfr_pow, args2, {NULL}, pow_cases, 0}, |
| 547 | {"powf", (funcptr)mpfr_pow, args2f, {NULL}, pow_cases_float, 0}, |
| 548 | /* |
| 549 | * Trig range reduction. We are able to test this for all |
| 550 | * finite values, but will only bother for things between 2^-3 |
| 551 | * and 2^+52. |
| 552 | */ |
| 553 | {"rred", (funcptr)test_rred, rred, {NULL}, rred_cases, 0}, |
| 554 | {"rredf", (funcptr)test_rred, rredf, {NULL}, rred_cases_float, 0}, |
| 555 | /* |
| 556 | * Square and cube root. |
| 557 | */ |
| 558 | {"sqrt", (funcptr)mpfr_sqrt, args1, {NULL}, log_cases, 0}, |
| 559 | {"sqrtf", (funcptr)mpfr_sqrt, args1f, {NULL}, log_cases_float, 0}, |
| 560 | {"cbrt", (funcptr)mpfr_cbrt, args1, {NULL}, log_cases, 0}, |
| 561 | {"cbrtf", (funcptr)mpfr_cbrt, args1f, {NULL}, log_cases_float, 0}, |
| 562 | {"hypot", (funcptr)mpfr_hypot, args2, {NULL}, atan2_cases, 0}, |
| 563 | {"hypotf", (funcptr)mpfr_hypot, args2f, {NULL}, atan2_cases_float, 0}, |
| 564 | /* |
| 565 | * Seminumerical functions. |
| 566 | */ |
| 567 | {"ceil", (funcptr)test_ceil, semi1, {NULL}, cases_semi1}, |
| 568 | {"ceilf", (funcptr)test_ceilf, semi1f, {NULL}, cases_semi1_float}, |
| 569 | {"floor", (funcptr)test_floor, semi1, {NULL}, cases_semi1}, |
| 570 | {"floorf", (funcptr)test_floorf, semi1f, {NULL}, cases_semi1_float}, |
| 571 | {"fmod", (funcptr)test_fmod, semi2, {NULL}, cases_semi2}, |
| 572 | {"fmodf", (funcptr)test_fmodf, semi2f, {NULL}, cases_semi2_float}, |
| 573 | {"ldexp", (funcptr)test_ldexp, t_ldexp, {NULL}, cases_ldexp}, |
| 574 | {"ldexpf", (funcptr)test_ldexpf, t_ldexpf, {NULL}, cases_ldexp_float}, |
| 575 | {"frexp", (funcptr)test_frexp, t_frexp, {NULL}, cases_semi1}, |
| 576 | {"frexpf", (funcptr)test_frexpf, t_frexpf, {NULL}, cases_semi1_float}, |
| 577 | {"modf", (funcptr)test_modf, t_modf, {NULL}, cases_semi1}, |
| 578 | {"modff", (funcptr)test_modff, t_modff, {NULL}, cases_semi1_float}, |
| 579 | |
| 580 | /* |
| 581 | * Classification and more semi-numericals |
| 582 | */ |
| 583 | {"copysign", (funcptr)test_copysign, semi2, {NULL}, cases_semi2}, |
| 584 | {"copysignf", (funcptr)test_copysignf, semi2f, {NULL}, cases_semi2_float}, |
| 585 | {"isfinite", (funcptr)test_isfinite, classify, {NULL}, cases_uniform, 0, 0x7fffffff}, |
| 586 | {"isfinitef", (funcptr)test_isfinitef, classifyf, {NULL}, cases_uniform_float, 0, 0x7fffffff}, |
| 587 | {"isinf", (funcptr)test_isinf, classify, {NULL}, cases_uniform, 0, 0x7fffffff}, |
| 588 | {"isinff", (funcptr)test_isinff, classifyf, {NULL}, cases_uniform_float, 0, 0x7fffffff}, |
| 589 | {"isnan", (funcptr)test_isnan, classify, {NULL}, cases_uniform, 0, 0x7fffffff}, |
| 590 | {"isnanf", (funcptr)test_isnanf, classifyf, {NULL}, cases_uniform_float, 0, 0x7fffffff}, |
| 591 | {"isnormal", (funcptr)test_isnormal, classify, {NULL}, cases_uniform, 0, 0x7fffffff}, |
| 592 | {"isnormalf", (funcptr)test_isnormalf, classifyf, {NULL}, cases_uniform_float, 0, 0x7fffffff}, |
| 593 | {"signbit", (funcptr)test_signbit, classify, {NULL}, cases_uniform, 0, 0x7fffffff}, |
| 594 | {"signbitf", (funcptr)test_signbitf, classifyf, {NULL}, cases_uniform_float, 0, 0x7fffffff}, |
| 595 | {"fpclassify", (funcptr)test_fpclassify, classify, {NULL}, cases_uniform, 0, 0x7fffffff}, |
| 596 | {"fpclassifyf", (funcptr)test_fpclassifyf, classifyf, {NULL}, cases_uniform_float, 0, 0x7fffffff}, |
| 597 | /* |
| 598 | * Comparisons |
| 599 | */ |
| 600 | {"isgreater", (funcptr)test_isgreater, compare, {NULL}, cases_uniform, 0, 0x7fffffff}, |
| 601 | {"isgreaterequal", (funcptr)test_isgreaterequal, compare, {NULL}, cases_uniform, 0, 0x7fffffff}, |
| 602 | {"isless", (funcptr)test_isless, compare, {NULL}, cases_uniform, 0, 0x7fffffff}, |
| 603 | {"islessequal", (funcptr)test_islessequal, compare, {NULL}, cases_uniform, 0, 0x7fffffff}, |
| 604 | {"islessgreater", (funcptr)test_islessgreater, compare, {NULL}, cases_uniform, 0, 0x7fffffff}, |
| 605 | {"isunordered", (funcptr)test_isunordered, compare, {NULL}, cases_uniform, 0, 0x7fffffff}, |
| 606 | |
| 607 | {"isgreaterf", (funcptr)test_isgreaterf, comparef, {NULL}, cases_uniform_float, 0, 0x7fffffff}, |
| 608 | {"isgreaterequalf", (funcptr)test_isgreaterequalf, comparef, {NULL}, cases_uniform_float, 0, 0x7fffffff}, |
| 609 | {"islessf", (funcptr)test_islessf, comparef, {NULL}, cases_uniform_float, 0, 0x7fffffff}, |
| 610 | {"islessequalf", (funcptr)test_islessequalf, comparef, {NULL}, cases_uniform_float, 0, 0x7fffffff}, |
| 611 | {"islessgreaterf", (funcptr)test_islessgreaterf, comparef, {NULL}, cases_uniform_float, 0, 0x7fffffff}, |
| 612 | {"isunorderedf", (funcptr)test_isunorderedf, comparef, {NULL}, cases_uniform_float, 0, 0x7fffffff}, |
| 613 | |
| 614 | /* |
| 615 | * Inverse Hyperbolic functions |
| 616 | */ |
| 617 | {"atanh", (funcptr)mpfr_atanh, args1, {NULL}, cases_uniform, 0x3e400000, 0x3fefffff}, |
| 618 | {"asinh", (funcptr)mpfr_asinh, args1, {NULL}, cases_uniform, 0x3e400000, 0x3fefffff}, |
| 619 | {"acosh", (funcptr)mpfr_acosh, args1, {NULL}, cases_uniform_positive, 0x3ff00000, 0x7fefffff}, |
| 620 | |
| 621 | {"atanhf", (funcptr)mpfr_atanh, args1f, {NULL}, cases_uniform_float, 0x32000000, 0x3f7fffff}, |
| 622 | {"asinhf", (funcptr)mpfr_asinh, args1f, {NULL}, cases_uniform_float, 0x32000000, 0x3f7fffff}, |
| 623 | {"acoshf", (funcptr)mpfr_acosh, args1f, {NULL}, cases_uniform_float_positive, 0x3f800000, 0x7f800000}, |
| 624 | |
| 625 | /* |
| 626 | * Everything else (sitting in a section down here at the bottom |
| 627 | * because historically they were not tested because we didn't |
| 628 | * have reference implementations for them) |
| 629 | */ |
| 630 | {"csin", (funcptr)mpc_sin, args1c, {NULL}, complex_cases_uniform, 0x3f000000, 0x40300000}, |
| 631 | {"csinf", (funcptr)mpc_sin, args1fc, {NULL}, complex_cases_uniform_float, 0x38000000, 0x41800000}, |
| 632 | {"ccos", (funcptr)mpc_cos, args1c, {NULL}, complex_cases_uniform, 0x3f000000, 0x40300000}, |
| 633 | {"ccosf", (funcptr)mpc_cos, args1fc, {NULL}, complex_cases_uniform_float, 0x38000000, 0x41800000}, |
| 634 | {"ctan", (funcptr)mpc_tan, args1c, {NULL}, complex_cases_uniform, 0x3f000000, 0x40300000}, |
| 635 | {"ctanf", (funcptr)mpc_tan, args1fc, {NULL}, complex_cases_uniform_float, 0x38000000, 0x41800000}, |
| 636 | |
| 637 | {"casin", (funcptr)mpc_asin, args1c, {NULL}, complex_cases_uniform, 0x3f000000, 0x40300000}, |
| 638 | {"casinf", (funcptr)mpc_asin, args1fc, {NULL}, complex_cases_uniform_float, 0x38000000, 0x41800000}, |
| 639 | {"cacos", (funcptr)mpc_acos, args1c, {NULL}, complex_cases_uniform, 0x3f000000, 0x40300000}, |
| 640 | {"cacosf", (funcptr)mpc_acos, args1fc, {NULL}, complex_cases_uniform_float, 0x38000000, 0x41800000}, |
| 641 | {"catan", (funcptr)mpc_atan, args1c, {NULL}, complex_cases_uniform, 0x3f000000, 0x40300000}, |
| 642 | {"catanf", (funcptr)mpc_atan, args1fc, {NULL}, complex_cases_uniform_float, 0x38000000, 0x41800000}, |
| 643 | |
| 644 | {"csinh", (funcptr)mpc_sinh, args1c, {NULL}, complex_cases_uniform, 0x3f000000, 0x40300000}, |
| 645 | {"csinhf", (funcptr)mpc_sinh, args1fc, {NULL}, complex_cases_uniform_float, 0x38000000, 0x41800000}, |
| 646 | {"ccosh", (funcptr)mpc_cosh, args1c, {NULL}, complex_cases_uniform, 0x3f000000, 0x40300000}, |
| 647 | {"ccoshf", (funcptr)mpc_cosh, args1fc, {NULL}, complex_cases_uniform_float, 0x38000000, 0x41800000}, |
| 648 | {"ctanh", (funcptr)mpc_tanh, args1c, {NULL}, complex_cases_uniform, 0x3f000000, 0x40300000}, |
| 649 | {"ctanhf", (funcptr)mpc_tanh, args1fc, {NULL}, complex_cases_uniform_float, 0x38000000, 0x41800000}, |
| 650 | |
| 651 | {"casinh", (funcptr)mpc_asinh, args1c, {NULL}, complex_cases_uniform, 0x3f000000, 0x40300000}, |
| 652 | {"casinhf", (funcptr)mpc_asinh, args1fc, {NULL}, complex_cases_uniform_float, 0x38000000, 0x41800000}, |
| 653 | {"cacosh", (funcptr)mpc_acosh, args1c, {NULL}, complex_cases_uniform, 0x3f000000, 0x40300000}, |
| 654 | {"cacoshf", (funcptr)mpc_acosh, args1fc, {NULL}, complex_cases_uniform_float, 0x38000000, 0x41800000}, |
| 655 | {"catanh", (funcptr)mpc_atanh, args1c, {NULL}, complex_cases_uniform, 0x3f000000, 0x40300000}, |
| 656 | {"catanhf", (funcptr)mpc_atanh, args1fc, {NULL}, complex_cases_uniform_float, 0x38000000, 0x41800000}, |
| 657 | |
| 658 | {"cexp", (funcptr)mpc_exp, args1c, {NULL}, complex_cases_uniform, 0x3c900000, 0x40862000}, |
| 659 | {"cpow", (funcptr)test_cpow, args2c, {NULL}, complex_pow_cases, 0x3fc00000, 0x40000000}, |
| 660 | {"clog", (funcptr)mpc_log, args1c, {NULL}, complex_log_cases, 0, 0}, |
| 661 | {"csqrt", (funcptr)mpc_sqrt, args1c, {NULL}, complex_log_cases, 0, 0}, |
| 662 | |
| 663 | {"cexpf", (funcptr)mpc_exp, args1fc, {NULL}, complex_cases_uniform_float, 0x24800000, 0x42b00000}, |
| 664 | {"cpowf", (funcptr)test_cpow, args2fc, {NULL}, complex_pow_cases_float, 0x3e000000, 0x41000000}, |
| 665 | {"clogf", (funcptr)mpc_log, args1fc, {NULL}, complex_log_cases_float, 0, 0}, |
| 666 | {"csqrtf", (funcptr)mpc_sqrt, args1fc, {NULL}, complex_log_cases_float, 0, 0}, |
| 667 | |
| 668 | {"cdiv", (funcptr)mpc_div, args2c, {NULL}, complex_arithmetic_cases, 0, 0}, |
| 669 | {"cmul", (funcptr)mpc_mul, args2c, {NULL}, complex_arithmetic_cases, 0, 0}, |
| 670 | {"cadd", (funcptr)mpc_add, args2c, {NULL}, complex_arithmetic_cases, 0, 0}, |
| 671 | {"csub", (funcptr)mpc_sub, args2c, {NULL}, complex_arithmetic_cases, 0, 0}, |
| 672 | |
| 673 | {"cdivf", (funcptr)mpc_div, args2fc, {NULL}, complex_arithmetic_cases_float, 0, 0}, |
| 674 | {"cmulf", (funcptr)mpc_mul, args2fc, {NULL}, complex_arithmetic_cases_float, 0, 0}, |
| 675 | {"caddf", (funcptr)mpc_add, args2fc, {NULL}, complex_arithmetic_cases_float, 0, 0}, |
| 676 | {"csubf", (funcptr)mpc_sub, args2fc, {NULL}, complex_arithmetic_cases_float, 0, 0}, |
| 677 | |
| 678 | {"cabsf", (funcptr)mpc_abs, args1fcr, {NULL}, complex_arithmetic_cases_float, 0, 0}, |
| 679 | {"cabs", (funcptr)mpc_abs, args1cr, {NULL}, complex_arithmetic_cases, 0, 0}, |
| 680 | {"cargf", (funcptr)mpc_arg, args1fcr, {NULL}, complex_arithmetic_cases_float, 0, 0}, |
| 681 | {"carg", (funcptr)mpc_arg, args1cr, {NULL}, complex_arithmetic_cases, 0, 0}, |
| 682 | {"cimagf", (funcptr)mpc_imag, args1fcr, {NULL}, complex_arithmetic_cases_float, 0, 0}, |
| 683 | {"cimag", (funcptr)mpc_imag, args1cr, {NULL}, complex_arithmetic_cases, 0, 0}, |
| 684 | {"conjf", (funcptr)mpc_conj, args1fc, {NULL}, complex_arithmetic_cases_float, 0, 0}, |
| 685 | {"conj", (funcptr)mpc_conj, args1c, {NULL}, complex_arithmetic_cases, 0, 0}, |
| 686 | {"cprojf", (funcptr)mpc_proj, args1fc, {NULL}, complex_arithmetic_cases_float, 0, 0}, |
| 687 | {"cproj", (funcptr)mpc_proj, args1c, {NULL}, complex_arithmetic_cases, 0, 0}, |
| 688 | {"crealf", (funcptr)mpc_real, args1fcr, {NULL}, complex_arithmetic_cases_float, 0, 0}, |
| 689 | {"creal", (funcptr)mpc_real, args1cr, {NULL}, complex_arithmetic_cases, 0, 0}, |
| 690 | {"erfcf", (funcptr)mpfr_erfc, args1f, {NULL}, cases_biased_float, 0x1e800000, 0x41000000}, |
| 691 | {"erfc", (funcptr)mpfr_erfc, args1, {NULL}, cases_biased, 0x3bd00000, 0x403c0000}, |
| 692 | {"erff", (funcptr)mpfr_erf, args1f, {NULL}, cases_biased_float, 0x03800000, 0x40700000}, |
| 693 | {"erf", (funcptr)mpfr_erf, args1, {NULL}, cases_biased, 0x00800000, 0x40200000}, |
| 694 | {"exp2f", (funcptr)mpfr_exp2, args1f, {NULL}, cases_uniform_float, 0x33800000, 0x43c00000}, |
| 695 | {"exp2", (funcptr)mpfr_exp2, args1, {NULL}, cases_uniform, 0x3ca00000, 0x40a00000}, |
| 696 | {"expm1f", (funcptr)mpfr_expm1, args1f, {NULL}, cases_uniform_float, 0x33000000, 0x43800000}, |
| 697 | {"expm1", (funcptr)mpfr_expm1, args1, {NULL}, cases_uniform, 0x3c900000, 0x409c0000}, |
| 698 | {"fmaxf", (funcptr)mpfr_max, args2f, {NULL}, minmax_cases_float, 0, 0x7f7fffff}, |
| 699 | {"fmax", (funcptr)mpfr_max, args2, {NULL}, minmax_cases, 0, 0x7fefffff}, |
| 700 | {"fminf", (funcptr)mpfr_min, args2f, {NULL}, minmax_cases_float, 0, 0x7f7fffff}, |
| 701 | {"fmin", (funcptr)mpfr_min, args2, {NULL}, minmax_cases, 0, 0x7fefffff}, |
| 702 | {"lgammaf", (funcptr)test_lgamma, args1f, {NULL}, cases_uniform_float, 0x01800000, 0x7f800000}, |
| 703 | {"lgamma", (funcptr)test_lgamma, args1, {NULL}, cases_uniform, 0x00100000, 0x7ff00000}, |
| 704 | {"log1pf", (funcptr)mpfr_log1p, args1f, {NULL}, log1p_cases_float, 0, 0}, |
| 705 | {"log1p", (funcptr)mpfr_log1p, args1, {NULL}, log1p_cases, 0, 0}, |
| 706 | {"log2f", (funcptr)mpfr_log2, args1f, {NULL}, log_cases_float, 0, 0}, |
| 707 | {"log2", (funcptr)mpfr_log2, args1, {NULL}, log_cases, 0, 0}, |
| 708 | {"tgammaf", (funcptr)mpfr_gamma, args1f, {NULL}, cases_uniform_float, 0x2f800000, 0x43000000}, |
| 709 | {"tgamma", (funcptr)mpfr_gamma, args1, {NULL}, cases_uniform, 0x3c000000, 0x40800000}, |
| 710 | }; |
| 711 | |
| 712 | const int nfunctions = ( sizeof(functions)/sizeof(*functions) ); |
| 713 | |
| 714 | #define random_sign ( random_upto(1) ? 0x80000000 : 0 ) |
| 715 | |
| 716 | static int iszero(uint32 *x) { |
| 717 | return !((x[0] & 0x7FFFFFFF) || x[1]); |
| 718 | } |
| 719 | |
| 720 | |
| 721 | static void complex_log_cases(uint32 *out, uint32 param1, |
| 722 | uint32 param2) { |
| 723 | cases_uniform(out,0x00100000,0x7fefffff); |
| 724 | cases_uniform(out+2,0x00100000,0x7fefffff); |
| 725 | } |
| 726 | |
| 727 | |
| 728 | static void complex_log_cases_float(uint32 *out, uint32 param1, |
| 729 | uint32 param2) { |
| 730 | cases_uniform_float(out,0x00800000,0x7f7fffff); |
| 731 | cases_uniform_float(out+2,0x00800000,0x7f7fffff); |
| 732 | } |
| 733 | |
| 734 | static void complex_cases_biased(uint32 *out, uint32 lowbound, |
| 735 | uint32 highbound) { |
| 736 | cases_biased(out,lowbound,highbound); |
| 737 | cases_biased(out+2,lowbound,highbound); |
| 738 | } |
| 739 | |
| 740 | static void complex_cases_biased_float(uint32 *out, uint32 lowbound, |
| 741 | uint32 highbound) { |
| 742 | cases_biased_float(out,lowbound,highbound); |
| 743 | cases_biased_float(out+2,lowbound,highbound); |
| 744 | } |
| 745 | |
| 746 | static void complex_cases_uniform(uint32 *out, uint32 lowbound, |
| 747 | uint32 highbound) { |
| 748 | cases_uniform(out,lowbound,highbound); |
| 749 | cases_uniform(out+2,lowbound,highbound); |
| 750 | } |
| 751 | |
| 752 | static void complex_cases_uniform_float(uint32 *out, uint32 lowbound, |
| 753 | uint32 highbound) { |
| 754 | cases_uniform_float(out,lowbound,highbound); |
| 755 | cases_uniform(out+2,lowbound,highbound); |
| 756 | } |
| 757 | |
| 758 | static void complex_pow_cases(uint32 *out, uint32 lowbound, |
| 759 | uint32 highbound) { |
| 760 | /* |
| 761 | * Generating non-overflowing cases for complex pow: |
| 762 | * |
| 763 | * Our base has both parts within the range [1/2,2], and hence |
| 764 | * its magnitude is within [1/2,2*sqrt(2)]. The magnitude of its |
| 765 | * logarithm in base 2 is therefore at most the magnitude of |
| 766 | * (log2(2*sqrt(2)) + i*pi/log(2)), or in other words |
| 767 | * hypot(3/2,pi/log(2)) = 4.77. So the magnitude of the exponent |
| 768 | * input must be at most our output magnitude limit (as a power |
| 769 | * of two) divided by that. |
| 770 | * |
| 771 | * I also set the output magnitude limit a bit low, because we |
| 772 | * don't guarantee (and neither does glibc) to prevent internal |
| 773 | * overflow in cases where the output _magnitude_ overflows but |
| 774 | * scaling it back down by cos and sin of the argument brings it |
| 775 | * back in range. |
| 776 | */ |
| 777 | cases_uniform(out,0x3fe00000, 0x40000000); |
| 778 | cases_uniform(out+2,0x3fe00000, 0x40000000); |
| 779 | cases_uniform(out+4,0x3f800000, 0x40600000); |
| 780 | cases_uniform(out+6,0x3f800000, 0x40600000); |
| 781 | } |
| 782 | |
| 783 | static void complex_pow_cases_float(uint32 *out, uint32 lowbound, |
| 784 | uint32 highbound) { |
| 785 | /* |
| 786 | * Reasoning as above, though of course the detailed numbers are |
| 787 | * all different. |
| 788 | */ |
| 789 | cases_uniform_float(out,0x3f000000, 0x40000000); |
| 790 | cases_uniform_float(out+2,0x3f000000, 0x40000000); |
| 791 | cases_uniform_float(out+4,0x3d600000, 0x41900000); |
| 792 | cases_uniform_float(out+6,0x3d600000, 0x41900000); |
| 793 | } |
| 794 | |
| 795 | static void complex_arithmetic_cases(uint32 *out, uint32 lowbound, |
| 796 | uint32 highbound) { |
| 797 | cases_uniform(out,0,0x7fefffff); |
| 798 | cases_uniform(out+2,0,0x7fefffff); |
| 799 | cases_uniform(out+4,0,0x7fefffff); |
| 800 | cases_uniform(out+6,0,0x7fefffff); |
| 801 | } |
| 802 | |
| 803 | static void complex_arithmetic_cases_float(uint32 *out, uint32 lowbound, |
| 804 | uint32 highbound) { |
| 805 | cases_uniform_float(out,0,0x7f7fffff); |
| 806 | cases_uniform_float(out+2,0,0x7f7fffff); |
| 807 | cases_uniform_float(out+4,0,0x7f7fffff); |
| 808 | cases_uniform_float(out+6,0,0x7f7fffff); |
| 809 | } |
| 810 | |
| 811 | /* |
| 812 | * Included from fplib test suite, in a compact self-contained |
| 813 | * form. |
| 814 | */ |
| 815 | |
| 816 | void float32_case(uint32 *ret) { |
| 817 | int n, bits; |
| 818 | uint32 f; |
| 819 | static int premax, preptr; |
| 820 | static uint32 *specifics = NULL; |
| 821 | |
| 822 | if (!ret) { |
| 823 | if (specifics) |
| 824 | free(specifics); |
| 825 | specifics = NULL; |
| 826 | premax = preptr = 0; |
| 827 | return; |
| 828 | } |
| 829 | |
| 830 | if (!specifics) { |
| 831 | int exps[] = { |
| 832 | -127, -126, -125, -24, -4, -3, -2, -1, 0, 1, 2, 3, 4, |
| 833 | 24, 29, 30, 31, 32, 61, 62, 63, 64, 126, 127, 128 |
| 834 | }; |
| 835 | int sign, eptr; |
| 836 | uint32 se, j; |
| 837 | /* |
| 838 | * We want a cross product of: |
| 839 | * - each of two sign bits (2) |
| 840 | * - each of the above (unbiased) exponents (25) |
| 841 | * - the following list of fraction parts: |
| 842 | * * zero (1) |
| 843 | * * all bits (1) |
| 844 | * * one-bit-set (23) |
| 845 | * * one-bit-clear (23) |
| 846 | * * one-bit-and-above (20: 3 are duplicates) |
| 847 | * * one-bit-and-below (20: 3 are duplicates) |
| 848 | * (total 88) |
| 849 | * (total 4400) |
| 850 | */ |
| 851 | specifics = malloc(4400 * sizeof(*specifics)); |
| 852 | preptr = 0; |
| 853 | for (sign = 0; sign <= 1; sign++) { |
| 854 | for (eptr = 0; eptr < sizeof(exps)/sizeof(*exps); eptr++) { |
| 855 | se = (sign ? 0x80000000 : 0) | ((exps[eptr]+127) << 23); |
| 856 | /* |
| 857 | * Zero. |
| 858 | */ |
| 859 | specifics[preptr++] = se | 0; |
| 860 | /* |
| 861 | * All bits. |
| 862 | */ |
| 863 | specifics[preptr++] = se | 0x7FFFFF; |
| 864 | /* |
| 865 | * One-bit-set. |
| 866 | */ |
| 867 | for (j = 1; j && j <= 0x400000; j <<= 1) |
| 868 | specifics[preptr++] = se | j; |
| 869 | /* |
| 870 | * One-bit-clear. |
| 871 | */ |
| 872 | for (j = 1; j && j <= 0x400000; j <<= 1) |
| 873 | specifics[preptr++] = se | (0x7FFFFF ^ j); |
| 874 | /* |
| 875 | * One-bit-and-everything-below. |
| 876 | */ |
| 877 | for (j = 2; j && j <= 0x100000; j <<= 1) |
| 878 | specifics[preptr++] = se | (2*j-1); |
| 879 | /* |
| 880 | * One-bit-and-everything-above. |
| 881 | */ |
| 882 | for (j = 4; j && j <= 0x200000; j <<= 1) |
| 883 | specifics[preptr++] = se | (0x7FFFFF ^ (j-1)); |
| 884 | /* |
| 885 | * Done. |
| 886 | */ |
| 887 | } |
| 888 | } |
| 889 | assert(preptr == 4400); |
| 890 | premax = preptr; |
| 891 | } |
| 892 | |
| 893 | /* |
| 894 | * Decide whether to return a pre or a random case. |
| 895 | */ |
| 896 | n = random32() % (premax+1); |
| 897 | if (n < preptr) { |
| 898 | /* |
| 899 | * Return pre[n]. |
| 900 | */ |
| 901 | uint32 t; |
| 902 | t = specifics[n]; |
| 903 | specifics[n] = specifics[preptr-1]; |
| 904 | specifics[preptr-1] = t; /* (not really needed) */ |
| 905 | preptr--; |
| 906 | *ret = t; |
| 907 | } else { |
| 908 | /* |
| 909 | * Random case. |
| 910 | * Sign and exponent: |
| 911 | * - FIXME |
| 912 | * Significand: |
| 913 | * - with prob 1/5, a totally random bit pattern |
| 914 | * - with prob 1/5, all 1s down to some point and then random |
| 915 | * - with prob 1/5, all 1s up to some point and then random |
| 916 | * - with prob 1/5, all 0s down to some point and then random |
| 917 | * - with prob 1/5, all 0s up to some point and then random |
| 918 | */ |
| 919 | n = random32() % 5; |
| 920 | f = random32(); /* some random bits */ |
| 921 | bits = random32() % 22 + 1; /* 1-22 */ |
| 922 | switch (n) { |
| 923 | case 0: |
| 924 | break; /* leave f alone */ |
| 925 | case 1: |
| 926 | f |= (1<<bits)-1; |
| 927 | break; |
| 928 | case 2: |
| 929 | f &= ~((1<<bits)-1); |
| 930 | break; |
| 931 | case 3: |
| 932 | f |= ~((1<<bits)-1); |
| 933 | break; |
| 934 | case 4: |
| 935 | f &= (1<<bits)-1; |
| 936 | break; |
| 937 | } |
| 938 | f &= 0x7FFFFF; |
| 939 | f |= (random32() & 0xFF800000);/* FIXME - do better */ |
| 940 | *ret = f; |
| 941 | } |
| 942 | } |
| 943 | static void float64_case(uint32 *ret) { |
| 944 | int n, bits; |
| 945 | uint32 f, g; |
| 946 | static int premax, preptr; |
| 947 | static uint32 (*specifics)[2] = NULL; |
| 948 | |
| 949 | if (!ret) { |
| 950 | if (specifics) |
| 951 | free(specifics); |
| 952 | specifics = NULL; |
| 953 | premax = preptr = 0; |
| 954 | return; |
| 955 | } |
| 956 | |
| 957 | if (!specifics) { |
| 958 | int exps[] = { |
| 959 | -1023, -1022, -1021, -129, -128, -127, -126, -53, -4, -3, -2, |
| 960 | -1, 0, 1, 2, 3, 4, 29, 30, 31, 32, 53, 61, 62, 63, 64, 127, |
| 961 | 128, 129, 1022, 1023, 1024 |
| 962 | }; |
| 963 | int sign, eptr; |
| 964 | uint32 se, j; |
| 965 | /* |
| 966 | * We want a cross product of: |
| 967 | * - each of two sign bits (2) |
| 968 | * - each of the above (unbiased) exponents (32) |
| 969 | * - the following list of fraction parts: |
| 970 | * * zero (1) |
| 971 | * * all bits (1) |
| 972 | * * one-bit-set (52) |
| 973 | * * one-bit-clear (52) |
| 974 | * * one-bit-and-above (49: 3 are duplicates) |
| 975 | * * one-bit-and-below (49: 3 are duplicates) |
| 976 | * (total 204) |
| 977 | * (total 13056) |
| 978 | */ |
| 979 | specifics = malloc(13056 * sizeof(*specifics)); |
| 980 | preptr = 0; |
| 981 | for (sign = 0; sign <= 1; sign++) { |
| 982 | for (eptr = 0; eptr < sizeof(exps)/sizeof(*exps); eptr++) { |
| 983 | se = (sign ? 0x80000000 : 0) | ((exps[eptr]+1023) << 20); |
| 984 | /* |
| 985 | * Zero. |
| 986 | */ |
| 987 | specifics[preptr][0] = 0; |
| 988 | specifics[preptr][1] = 0; |
| 989 | specifics[preptr++][0] |= se; |
| 990 | /* |
| 991 | * All bits. |
| 992 | */ |
| 993 | specifics[preptr][0] = 0xFFFFF; |
| 994 | specifics[preptr][1] = ~0; |
| 995 | specifics[preptr++][0] |= se; |
| 996 | /* |
| 997 | * One-bit-set. |
| 998 | */ |
| 999 | for (j = 1; j && j <= 0x80000000; j <<= 1) { |
| 1000 | specifics[preptr][0] = 0; |
| 1001 | specifics[preptr][1] = j; |
| 1002 | specifics[preptr++][0] |= se; |
| 1003 | if (j & 0xFFFFF) { |
| 1004 | specifics[preptr][0] = j; |
| 1005 | specifics[preptr][1] = 0; |
| 1006 | specifics[preptr++][0] |= se; |
| 1007 | } |
| 1008 | } |
| 1009 | /* |
| 1010 | * One-bit-clear. |
| 1011 | */ |
| 1012 | for (j = 1; j && j <= 0x80000000; j <<= 1) { |
| 1013 | specifics[preptr][0] = 0xFFFFF; |
| 1014 | specifics[preptr][1] = ~j; |
| 1015 | specifics[preptr++][0] |= se; |
| 1016 | if (j & 0xFFFFF) { |
| 1017 | specifics[preptr][0] = 0xFFFFF ^ j; |
| 1018 | specifics[preptr][1] = ~0; |
| 1019 | specifics[preptr++][0] |= se; |
| 1020 | } |
| 1021 | } |
| 1022 | /* |
| 1023 | * One-bit-and-everything-below. |
| 1024 | */ |
| 1025 | for (j = 2; j && j <= 0x80000000; j <<= 1) { |
| 1026 | specifics[preptr][0] = 0; |
| 1027 | specifics[preptr][1] = 2*j-1; |
| 1028 | specifics[preptr++][0] |= se; |
| 1029 | } |
| 1030 | for (j = 1; j && j <= 0x20000; j <<= 1) { |
| 1031 | specifics[preptr][0] = 2*j-1; |
| 1032 | specifics[preptr][1] = ~0; |
| 1033 | specifics[preptr++][0] |= se; |
| 1034 | } |
| 1035 | /* |
| 1036 | * One-bit-and-everything-above. |
| 1037 | */ |
| 1038 | for (j = 4; j && j <= 0x80000000; j <<= 1) { |
| 1039 | specifics[preptr][0] = 0xFFFFF; |
| 1040 | specifics[preptr][1] = ~(j-1); |
| 1041 | specifics[preptr++][0] |= se; |
| 1042 | } |
| 1043 | for (j = 1; j && j <= 0x40000; j <<= 1) { |
| 1044 | specifics[preptr][0] = 0xFFFFF ^ (j-1); |
| 1045 | specifics[preptr][1] = 0; |
| 1046 | specifics[preptr++][0] |= se; |
| 1047 | } |
| 1048 | /* |
| 1049 | * Done. |
| 1050 | */ |
| 1051 | } |
| 1052 | } |
| 1053 | assert(preptr == 13056); |
| 1054 | premax = preptr; |
| 1055 | } |
| 1056 | |
| 1057 | /* |
| 1058 | * Decide whether to return a pre or a random case. |
| 1059 | */ |
| 1060 | n = (uint32) random32() % (uint32) (premax+1); |
| 1061 | if (n < preptr) { |
| 1062 | /* |
| 1063 | * Return pre[n]. |
| 1064 | */ |
| 1065 | uint32 t; |
| 1066 | t = specifics[n][0]; |
| 1067 | specifics[n][0] = specifics[preptr-1][0]; |
| 1068 | specifics[preptr-1][0] = t; /* (not really needed) */ |
| 1069 | ret[0] = t; |
| 1070 | t = specifics[n][1]; |
| 1071 | specifics[n][1] = specifics[preptr-1][1]; |
| 1072 | specifics[preptr-1][1] = t; /* (not really needed) */ |
| 1073 | ret[1] = t; |
| 1074 | preptr--; |
| 1075 | } else { |
| 1076 | /* |
| 1077 | * Random case. |
| 1078 | * Sign and exponent: |
| 1079 | * - FIXME |
| 1080 | * Significand: |
| 1081 | * - with prob 1/5, a totally random bit pattern |
| 1082 | * - with prob 1/5, all 1s down to some point and then random |
| 1083 | * - with prob 1/5, all 1s up to some point and then random |
| 1084 | * - with prob 1/5, all 0s down to some point and then random |
| 1085 | * - with prob 1/5, all 0s up to some point and then random |
| 1086 | */ |
| 1087 | n = random32() % 5; |
| 1088 | f = random32(); /* some random bits */ |
| 1089 | g = random32(); /* some random bits */ |
| 1090 | bits = random32() % 51 + 1; /* 1-51 */ |
| 1091 | switch (n) { |
| 1092 | case 0: |
| 1093 | break; /* leave f alone */ |
| 1094 | case 1: |
| 1095 | if (bits <= 32) |
| 1096 | f |= (1<<bits)-1; |
| 1097 | else { |
| 1098 | bits -= 32; |
| 1099 | g |= (1<<bits)-1; |
| 1100 | f = ~0; |
| 1101 | } |
| 1102 | break; |
| 1103 | case 2: |
| 1104 | if (bits <= 32) |
| 1105 | f &= ~((1<<bits)-1); |
| 1106 | else { |
| 1107 | bits -= 32; |
| 1108 | g &= ~((1<<bits)-1); |
| 1109 | f = 0; |
| 1110 | } |
| 1111 | break; |
| 1112 | case 3: |
| 1113 | if (bits <= 32) |
| 1114 | g &= (1<<bits)-1; |
| 1115 | else { |
| 1116 | bits -= 32; |
| 1117 | f &= (1<<bits)-1; |
| 1118 | g = 0; |
| 1119 | } |
| 1120 | break; |
| 1121 | case 4: |
| 1122 | if (bits <= 32) |
| 1123 | g |= ~((1<<bits)-1); |
| 1124 | else { |
| 1125 | bits -= 32; |
| 1126 | f |= ~((1<<bits)-1); |
| 1127 | g = ~0; |
| 1128 | } |
| 1129 | break; |
| 1130 | } |
| 1131 | g &= 0xFFFFF; |
| 1132 | g |= (random32() & 0xFFF00000);/* FIXME - do better */ |
| 1133 | ret[0] = g; |
| 1134 | ret[1] = f; |
| 1135 | } |
| 1136 | } |
| 1137 | |
| 1138 | static void cases_biased(uint32 *out, uint32 lowbound, |
| 1139 | uint32 highbound) { |
| 1140 | do { |
| 1141 | out[0] = highbound - random_upto_biased(highbound-lowbound, 8); |
| 1142 | out[1] = random_upto(0xFFFFFFFF); |
| 1143 | out[0] |= random_sign; |
| 1144 | } while (iszero(out)); /* rule out zero */ |
| 1145 | } |
| 1146 | |
| 1147 | static void cases_biased_positive(uint32 *out, uint32 lowbound, |
| 1148 | uint32 highbound) { |
| 1149 | do { |
| 1150 | out[0] = highbound - random_upto_biased(highbound-lowbound, 8); |
| 1151 | out[1] = random_upto(0xFFFFFFFF); |
| 1152 | } while (iszero(out)); /* rule out zero */ |
| 1153 | } |
| 1154 | |
| 1155 | static void cases_biased_float(uint32 *out, uint32 lowbound, |
| 1156 | uint32 highbound) { |
| 1157 | do { |
| 1158 | out[0] = highbound - random_upto_biased(highbound-lowbound, 8); |
| 1159 | out[1] = 0; |
| 1160 | out[0] |= random_sign; |
| 1161 | } while (iszero(out)); /* rule out zero */ |
| 1162 | } |
| 1163 | |
| 1164 | static void cases_semi1(uint32 *out, uint32 param1, |
| 1165 | uint32 param2) { |
| 1166 | float64_case(out); |
| 1167 | } |
| 1168 | |
| 1169 | static void cases_semi1_float(uint32 *out, uint32 param1, |
| 1170 | uint32 param2) { |
| 1171 | float32_case(out); |
| 1172 | } |
| 1173 | |
| 1174 | static void cases_semi2(uint32 *out, uint32 param1, |
| 1175 | uint32 param2) { |
| 1176 | float64_case(out); |
| 1177 | float64_case(out+2); |
| 1178 | } |
| 1179 | |
| 1180 | static void cases_semi2_float(uint32 *out, uint32 param1, |
| 1181 | uint32 param2) { |
| 1182 | float32_case(out); |
| 1183 | float32_case(out+2); |
| 1184 | } |
| 1185 | |
| 1186 | static void cases_ldexp(uint32 *out, uint32 param1, |
| 1187 | uint32 param2) { |
| 1188 | float64_case(out); |
| 1189 | out[2] = random_upto(2048)-1024; |
| 1190 | } |
| 1191 | |
| 1192 | static void cases_ldexp_float(uint32 *out, uint32 param1, |
| 1193 | uint32 param2) { |
| 1194 | float32_case(out); |
| 1195 | out[2] = random_upto(256)-128; |
| 1196 | } |
| 1197 | |
| 1198 | static void cases_uniform(uint32 *out, uint32 lowbound, |
| 1199 | uint32 highbound) { |
| 1200 | do { |
| 1201 | out[0] = highbound - random_upto(highbound-lowbound); |
| 1202 | out[1] = random_upto(0xFFFFFFFF); |
| 1203 | out[0] |= random_sign; |
| 1204 | } while (iszero(out)); /* rule out zero */ |
| 1205 | } |
| 1206 | static void cases_uniform_float(uint32 *out, uint32 lowbound, |
| 1207 | uint32 highbound) { |
| 1208 | do { |
| 1209 | out[0] = highbound - random_upto(highbound-lowbound); |
| 1210 | out[1] = 0; |
| 1211 | out[0] |= random_sign; |
| 1212 | } while (iszero(out)); /* rule out zero */ |
| 1213 | } |
| 1214 | |
| 1215 | static void cases_uniform_positive(uint32 *out, uint32 lowbound, |
| 1216 | uint32 highbound) { |
| 1217 | do { |
| 1218 | out[0] = highbound - random_upto(highbound-lowbound); |
| 1219 | out[1] = random_upto(0xFFFFFFFF); |
| 1220 | } while (iszero(out)); /* rule out zero */ |
| 1221 | } |
| 1222 | static void cases_uniform_float_positive(uint32 *out, uint32 lowbound, |
| 1223 | uint32 highbound) { |
| 1224 | do { |
| 1225 | out[0] = highbound - random_upto(highbound-lowbound); |
| 1226 | out[1] = 0; |
| 1227 | } while (iszero(out)); /* rule out zero */ |
| 1228 | } |
| 1229 | |
| 1230 | |
| 1231 | static void log_cases(uint32 *out, uint32 param1, |
| 1232 | uint32 param2) { |
| 1233 | do { |
| 1234 | out[0] = random_upto(0x7FEFFFFF); |
| 1235 | out[1] = random_upto(0xFFFFFFFF); |
| 1236 | } while (iszero(out)); /* rule out zero */ |
| 1237 | } |
| 1238 | |
| 1239 | static void log_cases_float(uint32 *out, uint32 param1, |
| 1240 | uint32 param2) { |
| 1241 | do { |
| 1242 | out[0] = random_upto(0x7F7FFFFF); |
| 1243 | out[1] = 0; |
| 1244 | } while (iszero(out)); /* rule out zero */ |
| 1245 | } |
| 1246 | |
| 1247 | static void log1p_cases(uint32 *out, uint32 param1, uint32 param2) |
| 1248 | { |
| 1249 | uint32 sign = random_sign; |
| 1250 | if (sign == 0) { |
| 1251 | cases_uniform_positive(out, 0x3c700000, 0x43400000); |
| 1252 | } else { |
| 1253 | cases_uniform_positive(out, 0x3c000000, 0x3ff00000); |
| 1254 | } |
| 1255 | out[0] |= sign; |
| 1256 | } |
| 1257 | |
| 1258 | static void log1p_cases_float(uint32 *out, uint32 param1, uint32 param2) |
| 1259 | { |
| 1260 | uint32 sign = random_sign; |
| 1261 | if (sign == 0) { |
| 1262 | cases_uniform_float_positive(out, 0x32000000, 0x4c000000); |
| 1263 | } else { |
| 1264 | cases_uniform_float_positive(out, 0x30000000, 0x3f800000); |
| 1265 | } |
| 1266 | out[0] |= sign; |
| 1267 | } |
| 1268 | |
| 1269 | static void minmax_cases(uint32 *out, uint32 param1, uint32 param2) |
| 1270 | { |
| 1271 | do { |
| 1272 | out[0] = random_upto(0x7FEFFFFF); |
| 1273 | out[1] = random_upto(0xFFFFFFFF); |
| 1274 | out[0] |= random_sign; |
| 1275 | out[2] = random_upto(0x7FEFFFFF); |
| 1276 | out[3] = random_upto(0xFFFFFFFF); |
| 1277 | out[2] |= random_sign; |
| 1278 | } while (iszero(out)); /* rule out zero */ |
| 1279 | } |
| 1280 | |
| 1281 | static void minmax_cases_float(uint32 *out, uint32 param1, uint32 param2) |
| 1282 | { |
| 1283 | do { |
| 1284 | out[0] = random_upto(0x7F7FFFFF); |
| 1285 | out[1] = 0; |
| 1286 | out[0] |= random_sign; |
| 1287 | out[2] = random_upto(0x7F7FFFFF); |
| 1288 | out[3] = 0; |
| 1289 | out[2] |= random_sign; |
| 1290 | } while (iszero(out)); /* rule out zero */ |
| 1291 | } |
| 1292 | |
| 1293 | static void rred_cases(uint32 *out, uint32 param1, |
| 1294 | uint32 param2) { |
| 1295 | do { |
| 1296 | out[0] = ((0x3fc00000 + random_upto(0x036fffff)) | |
| 1297 | (random_upto(1) << 31)); |
| 1298 | out[1] = random_upto(0xFFFFFFFF); |
| 1299 | } while (iszero(out)); /* rule out zero */ |
| 1300 | } |
| 1301 | |
| 1302 | static void rred_cases_float(uint32 *out, uint32 param1, |
| 1303 | uint32 param2) { |
| 1304 | do { |
| 1305 | out[0] = ((0x3e000000 + random_upto(0x0cffffff)) | |
| 1306 | (random_upto(1) << 31)); |
| 1307 | out[1] = 0; /* for iszero */ |
| 1308 | } while (iszero(out)); /* rule out zero */ |
| 1309 | } |
| 1310 | |
| 1311 | static void atan2_cases(uint32 *out, uint32 param1, |
| 1312 | uint32 param2) { |
| 1313 | do { |
| 1314 | int expdiff = random_upto(101)-51; |
| 1315 | int swap; |
| 1316 | if (expdiff < 0) { |
| 1317 | expdiff = -expdiff; |
| 1318 | swap = 2; |
| 1319 | } else |
| 1320 | swap = 0; |
| 1321 | out[swap ^ 0] = random_upto(0x7FEFFFFF-((expdiff+1)<<20)); |
| 1322 | out[swap ^ 2] = random_upto(((expdiff+1)<<20)-1) + out[swap ^ 0]; |
| 1323 | out[1] = random_upto(0xFFFFFFFF); |
| 1324 | out[3] = random_upto(0xFFFFFFFF); |
| 1325 | out[0] |= random_sign; |
| 1326 | out[2] |= random_sign; |
| 1327 | } while (iszero(out) || iszero(out+2));/* rule out zero */ |
| 1328 | } |
| 1329 | |
| 1330 | static void atan2_cases_float(uint32 *out, uint32 param1, |
| 1331 | uint32 param2) { |
| 1332 | do { |
| 1333 | int expdiff = random_upto(44)-22; |
| 1334 | int swap; |
| 1335 | if (expdiff < 0) { |
| 1336 | expdiff = -expdiff; |
| 1337 | swap = 2; |
| 1338 | } else |
| 1339 | swap = 0; |
| 1340 | out[swap ^ 0] = random_upto(0x7F7FFFFF-((expdiff+1)<<23)); |
| 1341 | out[swap ^ 2] = random_upto(((expdiff+1)<<23)-1) + out[swap ^ 0]; |
| 1342 | out[0] |= random_sign; |
| 1343 | out[2] |= random_sign; |
| 1344 | out[1] = out[3] = 0; /* for iszero */ |
| 1345 | } while (iszero(out) || iszero(out+2));/* rule out zero */ |
| 1346 | } |
| 1347 | |
| 1348 | static void pow_cases(uint32 *out, uint32 param1, |
| 1349 | uint32 param2) { |
| 1350 | /* |
| 1351 | * Pick an exponent e (-0x33 to +0x7FE) for x, and here's the |
| 1352 | * range of numbers we can use as y: |
| 1353 | * |
| 1354 | * For e < 0x3FE, the range is [-0x400/(0x3FE-e),+0x432/(0x3FE-e)] |
| 1355 | * For e > 0x3FF, the range is [-0x432/(e-0x3FF),+0x400/(e-0x3FF)] |
| 1356 | * |
| 1357 | * For e == 0x3FE or e == 0x3FF, the range gets infinite at one |
| 1358 | * end or the other, so we have to be cleverer: pick a number n |
| 1359 | * of useful bits in the mantissa (1 thru 52, so 1 must imply |
| 1360 | * 0x3ff00000.00000001 whereas 52 is anything at least as big |
| 1361 | * as 0x3ff80000.00000000; for e == 0x3fe, 1 necessarily means |
| 1362 | * 0x3fefffff.ffffffff and 52 is anything at most as big as |
| 1363 | * 0x3fe80000.00000000). Then, as it happens, a sensible |
| 1364 | * maximum power is 2^(63-n) for e == 0x3fe, and 2^(62-n) for |
| 1365 | * e == 0x3ff. |
| 1366 | * |
| 1367 | * We inevitably get some overflows in approximating the log |
| 1368 | * curves by these nasty step functions, but that's all right - |
| 1369 | * we do want _some_ overflows to be tested. |
| 1370 | * |
| 1371 | * Having got that, then, it's just a matter of inventing a |
| 1372 | * probability distribution for all of this. |
| 1373 | */ |
| 1374 | int e, n; |
| 1375 | uint32 dmin, dmax; |
| 1376 | const uint32 pmin = 0x3e100000; |
| 1377 | |
| 1378 | /* |
| 1379 | * Generate exponents in a slightly biased fashion. |
| 1380 | */ |
| 1381 | e = (random_upto(1) ? /* is exponent small or big? */ |
| 1382 | 0x3FE - random_upto_biased(0x431,2) : /* small */ |
| 1383 | 0x3FF + random_upto_biased(0x3FF,2)); /* big */ |
| 1384 | |
| 1385 | /* |
| 1386 | * Now split into cases. |
| 1387 | */ |
| 1388 | if (e < 0x3FE || e > 0x3FF) { |
| 1389 | uint32 imin, imax; |
| 1390 | if (e < 0x3FE) |
| 1391 | imin = 0x40000 / (0x3FE - e), imax = 0x43200 / (0x3FE - e); |
| 1392 | else |
| 1393 | imin = 0x43200 / (e - 0x3FF), imax = 0x40000 / (e - 0x3FF); |
| 1394 | /* Power range runs from -imin to imax. Now convert to doubles */ |
| 1395 | dmin = doubletop(imin, -8); |
| 1396 | dmax = doubletop(imax, -8); |
| 1397 | /* Compute the number of mantissa bits. */ |
| 1398 | n = (e > 0 ? 53 : 52+e); |
| 1399 | } else { |
| 1400 | /* Critical exponents. Generate a top bit index. */ |
| 1401 | n = 52 - random_upto_biased(51, 4); |
| 1402 | if (e == 0x3FE) |
| 1403 | dmax = 63 - n; |
| 1404 | else |
| 1405 | dmax = 62 - n; |
| 1406 | dmax = (dmax << 20) + 0x3FF00000; |
| 1407 | dmin = dmax; |
| 1408 | } |
| 1409 | /* Generate a mantissa. */ |
| 1410 | if (n <= 32) { |
| 1411 | out[0] = 0; |
| 1412 | out[1] = random_upto((1 << (n-1)) - 1) + (1 << (n-1)); |
| 1413 | } else if (n == 33) { |
| 1414 | out[0] = 1; |
| 1415 | out[1] = random_upto(0xFFFFFFFF); |
| 1416 | } else if (n > 33) { |
| 1417 | out[0] = random_upto((1 << (n-33)) - 1) + (1 << (n-33)); |
| 1418 | out[1] = random_upto(0xFFFFFFFF); |
| 1419 | } |
| 1420 | /* Negate the mantissa if e == 0x3FE. */ |
| 1421 | if (e == 0x3FE) { |
| 1422 | out[1] = -out[1]; |
| 1423 | out[0] = -out[0]; |
| 1424 | if (out[1]) out[0]--; |
| 1425 | } |
| 1426 | /* Put the exponent on. */ |
| 1427 | out[0] &= 0xFFFFF; |
| 1428 | out[0] |= ((e > 0 ? e : 0) << 20); |
| 1429 | /* Generate a power. Powers don't go below 2^-30. */ |
| 1430 | if (random_upto(1)) { |
| 1431 | /* Positive power */ |
| 1432 | out[2] = dmax - random_upto_biased(dmax-pmin, 10); |
| 1433 | } else { |
| 1434 | /* Negative power */ |
| 1435 | out[2] = (dmin - random_upto_biased(dmin-pmin, 10)) | 0x80000000; |
| 1436 | } |
| 1437 | out[3] = random_upto(0xFFFFFFFF); |
| 1438 | } |
| 1439 | static void pow_cases_float(uint32 *out, uint32 param1, |
| 1440 | uint32 param2) { |
| 1441 | /* |
| 1442 | * Pick an exponent e (-0x16 to +0xFE) for x, and here's the |
| 1443 | * range of numbers we can use as y: |
| 1444 | * |
| 1445 | * For e < 0x7E, the range is [-0x80/(0x7E-e),+0x95/(0x7E-e)] |
| 1446 | * For e > 0x7F, the range is [-0x95/(e-0x7F),+0x80/(e-0x7F)] |
| 1447 | * |
| 1448 | * For e == 0x7E or e == 0x7F, the range gets infinite at one |
| 1449 | * end or the other, so we have to be cleverer: pick a number n |
| 1450 | * of useful bits in the mantissa (1 thru 23, so 1 must imply |
| 1451 | * 0x3f800001 whereas 23 is anything at least as big as |
| 1452 | * 0x3fc00000; for e == 0x7e, 1 necessarily means 0x3f7fffff |
| 1453 | * and 23 is anything at most as big as 0x3f400000). Then, as |
| 1454 | * it happens, a sensible maximum power is 2^(31-n) for e == |
| 1455 | * 0x7e, and 2^(30-n) for e == 0x7f. |
| 1456 | * |
| 1457 | * We inevitably get some overflows in approximating the log |
| 1458 | * curves by these nasty step functions, but that's all right - |
| 1459 | * we do want _some_ overflows to be tested. |
| 1460 | * |
| 1461 | * Having got that, then, it's just a matter of inventing a |
| 1462 | * probability distribution for all of this. |
| 1463 | */ |
| 1464 | int e, n; |
| 1465 | uint32 dmin, dmax; |
| 1466 | const uint32 pmin = 0x38000000; |
| 1467 | |
| 1468 | /* |
| 1469 | * Generate exponents in a slightly biased fashion. |
| 1470 | */ |
| 1471 | e = (random_upto(1) ? /* is exponent small or big? */ |
| 1472 | 0x7E - random_upto_biased(0x94,2) : /* small */ |
| 1473 | 0x7F + random_upto_biased(0x7f,2)); /* big */ |
| 1474 | |
| 1475 | /* |
| 1476 | * Now split into cases. |
| 1477 | */ |
| 1478 | if (e < 0x7E || e > 0x7F) { |
| 1479 | uint32 imin, imax; |
| 1480 | if (e < 0x7E) |
| 1481 | imin = 0x8000 / (0x7e - e), imax = 0x9500 / (0x7e - e); |
| 1482 | else |
| 1483 | imin = 0x9500 / (e - 0x7f), imax = 0x8000 / (e - 0x7f); |
| 1484 | /* Power range runs from -imin to imax. Now convert to doubles */ |
| 1485 | dmin = floatval(imin, -8); |
| 1486 | dmax = floatval(imax, -8); |
| 1487 | /* Compute the number of mantissa bits. */ |
| 1488 | n = (e > 0 ? 24 : 23+e); |
| 1489 | } else { |
| 1490 | /* Critical exponents. Generate a top bit index. */ |
| 1491 | n = 23 - random_upto_biased(22, 4); |
| 1492 | if (e == 0x7E) |
| 1493 | dmax = 31 - n; |
| 1494 | else |
| 1495 | dmax = 30 - n; |
| 1496 | dmax = (dmax << 23) + 0x3F800000; |
| 1497 | dmin = dmax; |
| 1498 | } |
| 1499 | /* Generate a mantissa. */ |
| 1500 | out[0] = random_upto((1 << (n-1)) - 1) + (1 << (n-1)); |
| 1501 | out[1] = 0; |
| 1502 | /* Negate the mantissa if e == 0x7E. */ |
| 1503 | if (e == 0x7E) { |
| 1504 | out[0] = -out[0]; |
| 1505 | } |
| 1506 | /* Put the exponent on. */ |
| 1507 | out[0] &= 0x7FFFFF; |
| 1508 | out[0] |= ((e > 0 ? e : 0) << 23); |
| 1509 | /* Generate a power. Powers don't go below 2^-15. */ |
| 1510 | if (random_upto(1)) { |
| 1511 | /* Positive power */ |
| 1512 | out[2] = dmax - random_upto_biased(dmax-pmin, 10); |
| 1513 | } else { |
| 1514 | /* Negative power */ |
| 1515 | out[2] = (dmin - random_upto_biased(dmin-pmin, 10)) | 0x80000000; |
| 1516 | } |
| 1517 | out[3] = 0; |
| 1518 | } |
| 1519 | |
| 1520 | void vet_for_decline(Testable *fn, uint32 *args, uint32 *result, int got_errno_in) { |
| 1521 | int declined = 0; |
| 1522 | |
| 1523 | switch (fn->type) { |
| 1524 | case args1: |
| 1525 | case rred: |
| 1526 | case semi1: |
| 1527 | case t_frexp: |
| 1528 | case t_modf: |
| 1529 | case classify: |
| 1530 | case t_ldexp: |
| 1531 | declined |= lib_fo && is_dhard(args+0); |
| 1532 | break; |
| 1533 | case args1f: |
| 1534 | case rredf: |
| 1535 | case semi1f: |
| 1536 | case t_frexpf: |
| 1537 | case t_modff: |
| 1538 | case classifyf: |
| 1539 | declined |= lib_fo && is_shard(args+0); |
| 1540 | break; |
| 1541 | case args2: |
| 1542 | case semi2: |
| 1543 | case args1c: |
| 1544 | case args1cr: |
| 1545 | case compare: |
| 1546 | declined |= lib_fo && is_dhard(args+0); |
| 1547 | declined |= lib_fo && is_dhard(args+2); |
| 1548 | break; |
| 1549 | case args2f: |
| 1550 | case semi2f: |
| 1551 | case t_ldexpf: |
| 1552 | case comparef: |
| 1553 | case args1fc: |
| 1554 | case args1fcr: |
| 1555 | declined |= lib_fo && is_shard(args+0); |
| 1556 | declined |= lib_fo && is_shard(args+2); |
| 1557 | break; |
| 1558 | case args2c: |
| 1559 | declined |= lib_fo && is_dhard(args+0); |
| 1560 | declined |= lib_fo && is_dhard(args+2); |
| 1561 | declined |= lib_fo && is_dhard(args+4); |
| 1562 | declined |= lib_fo && is_dhard(args+6); |
| 1563 | break; |
| 1564 | case args2fc: |
| 1565 | declined |= lib_fo && is_shard(args+0); |
| 1566 | declined |= lib_fo && is_shard(args+2); |
| 1567 | declined |= lib_fo && is_shard(args+4); |
| 1568 | declined |= lib_fo && is_shard(args+6); |
| 1569 | break; |
| 1570 | } |
| 1571 | |
| 1572 | switch (fn->type) { |
| 1573 | case args1: /* return an extra-precise result */ |
| 1574 | case args2: |
| 1575 | case rred: |
| 1576 | case semi1: /* return a double result */ |
| 1577 | case semi2: |
| 1578 | case t_ldexp: |
| 1579 | case t_frexp: /* return double * int */ |
| 1580 | case args1cr: |
| 1581 | declined |= lib_fo && is_dhard(result); |
| 1582 | break; |
| 1583 | case args1f: |
| 1584 | case args2f: |
| 1585 | case rredf: |
| 1586 | case semi1f: |
| 1587 | case semi2f: |
| 1588 | case t_ldexpf: |
| 1589 | case args1fcr: |
| 1590 | declined |= lib_fo && is_shard(result); |
| 1591 | break; |
| 1592 | case t_modf: /* return double * double */ |
| 1593 | declined |= lib_fo && is_dhard(result+0); |
| 1594 | declined |= lib_fo && is_dhard(result+2); |
| 1595 | break; |
| 1596 | case t_modff: /* return float * float */ |
| 1597 | declined |= lib_fo && is_shard(result+2); |
| 1598 | /* fall through */ |
| 1599 | case t_frexpf: /* return float * int */ |
| 1600 | declined |= lib_fo && is_shard(result+0); |
| 1601 | break; |
| 1602 | case args1c: |
| 1603 | case args2c: |
| 1604 | declined |= lib_fo && is_dhard(result+0); |
| 1605 | declined |= lib_fo && is_dhard(result+4); |
| 1606 | break; |
| 1607 | case args1fc: |
| 1608 | case args2fc: |
| 1609 | declined |= lib_fo && is_shard(result+0); |
| 1610 | declined |= lib_fo && is_shard(result+4); |
| 1611 | break; |
| 1612 | } |
| 1613 | |
| 1614 | /* Expect basic arithmetic tests to be declined if the command |
| 1615 | * line said that would happen */ |
| 1616 | declined |= (lib_no_arith && (fn->func == (funcptr)mpc_add || |
| 1617 | fn->func == (funcptr)mpc_sub || |
| 1618 | fn->func == (funcptr)mpc_mul || |
| 1619 | fn->func == (funcptr)mpc_div)); |
| 1620 | |
| 1621 | if (!declined) { |
| 1622 | if (got_errno_in) |
| 1623 | ntests++; |
| 1624 | else |
| 1625 | ntests += 3; |
| 1626 | } |
| 1627 | } |
| 1628 | |
| 1629 | void docase(Testable *fn, uint32 *args) { |
| 1630 | uint32 result[8]; /* real part in first 4, imaginary part in last 4 */ |
| 1631 | char *errstr = NULL; |
| 1632 | mpfr_t a, b, r; |
| 1633 | mpc_t ac, bc, rc; |
| 1634 | int rejected, printextra; |
| 1635 | wrapperctx ctx; |
| 1636 | |
| 1637 | mpfr_init2(a, MPFR_PREC); |
| 1638 | mpfr_init2(b, MPFR_PREC); |
| 1639 | mpfr_init2(r, MPFR_PREC); |
| 1640 | mpc_init2(ac, MPFR_PREC); |
| 1641 | mpc_init2(bc, MPFR_PREC); |
| 1642 | mpc_init2(rc, MPFR_PREC); |
| 1643 | |
| 1644 | printf("func=%s", fn->name); |
| 1645 | |
| 1646 | rejected = 0; /* FIXME */ |
| 1647 | |
| 1648 | switch (fn->type) { |
| 1649 | case args1: |
| 1650 | case rred: |
| 1651 | case semi1: |
| 1652 | case t_frexp: |
| 1653 | case t_modf: |
| 1654 | case classify: |
| 1655 | printf(" op1=%08x.%08x", args[0], args[1]); |
| 1656 | break; |
| 1657 | case args1f: |
| 1658 | case rredf: |
| 1659 | case semi1f: |
| 1660 | case t_frexpf: |
| 1661 | case t_modff: |
| 1662 | case classifyf: |
| 1663 | printf(" op1=%08x", args[0]); |
| 1664 | break; |
| 1665 | case args2: |
| 1666 | case semi2: |
| 1667 | case compare: |
| 1668 | printf(" op1=%08x.%08x", args[0], args[1]); |
| 1669 | printf(" op2=%08x.%08x", args[2], args[3]); |
| 1670 | break; |
| 1671 | case args2f: |
| 1672 | case semi2f: |
| 1673 | case t_ldexpf: |
| 1674 | case comparef: |
| 1675 | printf(" op1=%08x", args[0]); |
| 1676 | printf(" op2=%08x", args[2]); |
| 1677 | break; |
| 1678 | case t_ldexp: |
| 1679 | printf(" op1=%08x.%08x", args[0], args[1]); |
| 1680 | printf(" op2=%08x", args[2]); |
| 1681 | break; |
| 1682 | case args1c: |
| 1683 | case args1cr: |
| 1684 | printf(" op1r=%08x.%08x", args[0], args[1]); |
| 1685 | printf(" op1i=%08x.%08x", args[2], args[3]); |
| 1686 | break; |
| 1687 | case args2c: |
| 1688 | printf(" op1r=%08x.%08x", args[0], args[1]); |
| 1689 | printf(" op1i=%08x.%08x", args[2], args[3]); |
| 1690 | printf(" op2r=%08x.%08x", args[4], args[5]); |
| 1691 | printf(" op2i=%08x.%08x", args[6], args[7]); |
| 1692 | break; |
| 1693 | case args1fc: |
| 1694 | case args1fcr: |
| 1695 | printf(" op1r=%08x", args[0]); |
| 1696 | printf(" op1i=%08x", args[2]); |
| 1697 | break; |
| 1698 | case args2fc: |
| 1699 | printf(" op1r=%08x", args[0]); |
| 1700 | printf(" op1i=%08x", args[2]); |
| 1701 | printf(" op2r=%08x", args[4]); |
| 1702 | printf(" op2i=%08x", args[6]); |
| 1703 | break; |
| 1704 | default: |
| 1705 | fprintf(stderr, "internal inconsistency?!\n"); |
| 1706 | abort(); |
| 1707 | } |
| 1708 | |
| 1709 | if (rejected == 2) { |
| 1710 | printf(" - test case rejected\n"); |
| 1711 | goto cleanup; |
| 1712 | } |
| 1713 | |
| 1714 | wrapper_init(&ctx); |
| 1715 | |
| 1716 | if (rejected == 0) { |
| 1717 | switch (fn->type) { |
| 1718 | case args1: |
| 1719 | set_mpfr_d(a, args[0], args[1]); |
| 1720 | wrapper_op_real(&ctx, a, 2, args); |
| 1721 | ((testfunc1)(fn->func))(r, a, GMP_RNDN); |
| 1722 | get_mpfr_d(r, &result[0], &result[1], &result[2]); |
| 1723 | wrapper_result_real(&ctx, r, 2, result); |
| 1724 | if (wrapper_run(&ctx, fn->wrappers)) |
| 1725 | get_mpfr_d(r, &result[0], &result[1], &result[2]); |
| 1726 | break; |
| 1727 | case args1cr: |
| 1728 | set_mpc_d(ac, args[0], args[1], args[2], args[3]); |
| 1729 | wrapper_op_complex(&ctx, ac, 2, args); |
| 1730 | ((testfunc1cr)(fn->func))(r, ac, GMP_RNDN); |
| 1731 | get_mpfr_d(r, &result[0], &result[1], &result[2]); |
| 1732 | wrapper_result_real(&ctx, r, 2, result); |
| 1733 | if (wrapper_run(&ctx, fn->wrappers)) |
| 1734 | get_mpfr_d(r, &result[0], &result[1], &result[2]); |
| 1735 | break; |
| 1736 | case args1f: |
| 1737 | set_mpfr_f(a, args[0]); |
| 1738 | wrapper_op_real(&ctx, a, 1, args); |
| 1739 | ((testfunc1)(fn->func))(r, a, GMP_RNDN); |
| 1740 | get_mpfr_f(r, &result[0], &result[1]); |
| 1741 | wrapper_result_real(&ctx, r, 1, result); |
| 1742 | if (wrapper_run(&ctx, fn->wrappers)) |
| 1743 | get_mpfr_f(r, &result[0], &result[1]); |
| 1744 | break; |
| 1745 | case args1fcr: |
| 1746 | set_mpc_f(ac, args[0], args[2]); |
| 1747 | wrapper_op_complex(&ctx, ac, 1, args); |
| 1748 | ((testfunc1cr)(fn->func))(r, ac, GMP_RNDN); |
| 1749 | get_mpfr_f(r, &result[0], &result[1]); |
| 1750 | wrapper_result_real(&ctx, r, 1, result); |
| 1751 | if (wrapper_run(&ctx, fn->wrappers)) |
| 1752 | get_mpfr_f(r, &result[0], &result[1]); |
| 1753 | break; |
| 1754 | case args2: |
| 1755 | set_mpfr_d(a, args[0], args[1]); |
| 1756 | wrapper_op_real(&ctx, a, 2, args); |
| 1757 | set_mpfr_d(b, args[2], args[3]); |
| 1758 | wrapper_op_real(&ctx, b, 2, args+2); |
| 1759 | ((testfunc2)(fn->func))(r, a, b, GMP_RNDN); |
| 1760 | get_mpfr_d(r, &result[0], &result[1], &result[2]); |
| 1761 | wrapper_result_real(&ctx, r, 2, result); |
| 1762 | if (wrapper_run(&ctx, fn->wrappers)) |
| 1763 | get_mpfr_d(r, &result[0], &result[1], &result[2]); |
| 1764 | break; |
| 1765 | case args2f: |
| 1766 | set_mpfr_f(a, args[0]); |
| 1767 | wrapper_op_real(&ctx, a, 1, args); |
| 1768 | set_mpfr_f(b, args[2]); |
| 1769 | wrapper_op_real(&ctx, b, 1, args+2); |
| 1770 | ((testfunc2)(fn->func))(r, a, b, GMP_RNDN); |
| 1771 | get_mpfr_f(r, &result[0], &result[1]); |
| 1772 | wrapper_result_real(&ctx, r, 1, result); |
| 1773 | if (wrapper_run(&ctx, fn->wrappers)) |
| 1774 | get_mpfr_f(r, &result[0], &result[1]); |
| 1775 | break; |
| 1776 | case rred: |
| 1777 | set_mpfr_d(a, args[0], args[1]); |
| 1778 | wrapper_op_real(&ctx, a, 2, args); |
| 1779 | ((testrred)(fn->func))(r, a, (int *)&result[3]); |
| 1780 | get_mpfr_d(r, &result[0], &result[1], &result[2]); |
| 1781 | wrapper_result_real(&ctx, r, 2, result); |
| 1782 | /* We never need to mess about with the integer auxiliary |
| 1783 | * output. */ |
| 1784 | if (wrapper_run(&ctx, fn->wrappers)) |
| 1785 | get_mpfr_d(r, &result[0], &result[1], &result[2]); |
| 1786 | break; |
| 1787 | case rredf: |
| 1788 | set_mpfr_f(a, args[0]); |
| 1789 | wrapper_op_real(&ctx, a, 1, args); |
| 1790 | ((testrred)(fn->func))(r, a, (int *)&result[3]); |
| 1791 | get_mpfr_f(r, &result[0], &result[1]); |
| 1792 | wrapper_result_real(&ctx, r, 1, result); |
| 1793 | /* We never need to mess about with the integer auxiliary |
| 1794 | * output. */ |
| 1795 | if (wrapper_run(&ctx, fn->wrappers)) |
| 1796 | get_mpfr_f(r, &result[0], &result[1]); |
| 1797 | break; |
| 1798 | case semi1: |
| 1799 | case semi1f: |
| 1800 | errstr = ((testsemi1)(fn->func))(args, result); |
| 1801 | break; |
| 1802 | case semi2: |
| 1803 | case compare: |
| 1804 | errstr = ((testsemi2)(fn->func))(args, args+2, result); |
| 1805 | break; |
| 1806 | case semi2f: |
| 1807 | case comparef: |
| 1808 | case t_ldexpf: |
| 1809 | errstr = ((testsemi2f)(fn->func))(args, args+2, result); |
| 1810 | break; |
| 1811 | case t_ldexp: |
| 1812 | errstr = ((testldexp)(fn->func))(args, args+2, result); |
| 1813 | break; |
| 1814 | case t_frexp: |
| 1815 | errstr = ((testfrexp)(fn->func))(args, result, result+2); |
| 1816 | break; |
| 1817 | case t_frexpf: |
| 1818 | errstr = ((testfrexp)(fn->func))(args, result, result+2); |
| 1819 | break; |
| 1820 | case t_modf: |
| 1821 | errstr = ((testmodf)(fn->func))(args, result, result+2); |
| 1822 | break; |
| 1823 | case t_modff: |
| 1824 | errstr = ((testmodf)(fn->func))(args, result, result+2); |
| 1825 | break; |
| 1826 | case classify: |
| 1827 | errstr = ((testclassify)(fn->func))(args, &result[0]); |
| 1828 | break; |
| 1829 | case classifyf: |
| 1830 | errstr = ((testclassifyf)(fn->func))(args, &result[0]); |
| 1831 | break; |
| 1832 | case args1c: |
| 1833 | set_mpc_d(ac, args[0], args[1], args[2], args[3]); |
| 1834 | wrapper_op_complex(&ctx, ac, 2, args); |
| 1835 | ((testfunc1c)(fn->func))(rc, ac, MPC_RNDNN); |
| 1836 | get_mpc_d(rc, &result[0], &result[1], &result[2], &result[4], &result[5], &result[6]); |
| 1837 | wrapper_result_complex(&ctx, rc, 2, result); |
| 1838 | if (wrapper_run(&ctx, fn->wrappers)) |
| 1839 | get_mpc_d(rc, &result[0], &result[1], &result[2], &result[4], &result[5], &result[6]); |
| 1840 | break; |
| 1841 | case args2c: |
| 1842 | set_mpc_d(ac, args[0], args[1], args[2], args[3]); |
| 1843 | wrapper_op_complex(&ctx, ac, 2, args); |
| 1844 | set_mpc_d(bc, args[4], args[5], args[6], args[7]); |
| 1845 | wrapper_op_complex(&ctx, bc, 2, args+4); |
| 1846 | ((testfunc2c)(fn->func))(rc, ac, bc, MPC_RNDNN); |
| 1847 | get_mpc_d(rc, &result[0], &result[1], &result[2], &result[4], &result[5], &result[6]); |
| 1848 | wrapper_result_complex(&ctx, rc, 2, result); |
| 1849 | if (wrapper_run(&ctx, fn->wrappers)) |
| 1850 | get_mpc_d(rc, &result[0], &result[1], &result[2], &result[4], &result[5], &result[6]); |
| 1851 | break; |
| 1852 | case args1fc: |
| 1853 | set_mpc_f(ac, args[0], args[2]); |
| 1854 | wrapper_op_complex(&ctx, ac, 1, args); |
| 1855 | ((testfunc1c)(fn->func))(rc, ac, MPC_RNDNN); |
| 1856 | get_mpc_f(rc, &result[0], &result[1], &result[4], &result[5]); |
| 1857 | wrapper_result_complex(&ctx, rc, 1, result); |
| 1858 | if (wrapper_run(&ctx, fn->wrappers)) |
| 1859 | get_mpc_f(rc, &result[0], &result[1], &result[4], &result[5]); |
| 1860 | break; |
| 1861 | case args2fc: |
| 1862 | set_mpc_f(ac, args[0], args[2]); |
| 1863 | wrapper_op_complex(&ctx, ac, 1, args); |
| 1864 | set_mpc_f(bc, args[4], args[6]); |
| 1865 | wrapper_op_complex(&ctx, bc, 1, args+4); |
| 1866 | ((testfunc2c)(fn->func))(rc, ac, bc, MPC_RNDNN); |
| 1867 | get_mpc_f(rc, &result[0], &result[1], &result[4], &result[5]); |
| 1868 | wrapper_result_complex(&ctx, rc, 1, result); |
| 1869 | if (wrapper_run(&ctx, fn->wrappers)) |
| 1870 | get_mpc_f(rc, &result[0], &result[1], &result[4], &result[5]); |
| 1871 | break; |
| 1872 | default: |
| 1873 | fprintf(stderr, "internal inconsistency?!\n"); |
| 1874 | abort(); |
| 1875 | } |
| 1876 | } |
| 1877 | |
| 1878 | switch (fn->type) { |
| 1879 | case args1: /* return an extra-precise result */ |
| 1880 | case args2: |
| 1881 | case args1cr: |
| 1882 | case rred: |
| 1883 | printextra = 1; |
| 1884 | if (rejected == 0) { |
| 1885 | errstr = NULL; |
| 1886 | if (!mpfr_zero_p(a)) { |
| 1887 | if ((result[0] & 0x7FFFFFFF) == 0 && result[1] == 0) { |
| 1888 | /* |
| 1889 | * If the output is +0 or -0 apart from the extra |
| 1890 | * precision in result[2], then there's a tricky |
| 1891 | * judgment call about what we require in the |
| 1892 | * output. If we output the extra bits and set |
| 1893 | * errstr="?underflow" then mathtest will tolerate |
| 1894 | * the function under test rounding down to zero |
| 1895 | * _or_ up to the minimum denormal; whereas if we |
| 1896 | * suppress the extra bits and set |
| 1897 | * errstr="underflow", then mathtest will enforce |
| 1898 | * that the function really does underflow to zero. |
| 1899 | * |
| 1900 | * But where to draw the line? It seems clear to |
| 1901 | * me that numbers along the lines of |
| 1902 | * 00000000.00000000.7ff should be treated |
| 1903 | * similarly to 00000000.00000000.801, but on the |
| 1904 | * other hand, we must surely be prepared to |
| 1905 | * enforce a genuine underflow-to-zero in _some_ |
| 1906 | * case where the true mathematical output is |
| 1907 | * nonzero but absurdly tiny. |
| 1908 | * |
| 1909 | * I think a reasonable place to draw the |
| 1910 | * distinction is at 00000000.00000000.400, i.e. |
| 1911 | * one quarter of the minimum positive denormal. |
| 1912 | * If a value less than that rounds up to the |
| 1913 | * minimum denormal, that must mean the function |
| 1914 | * under test has managed to make an error of an |
| 1915 | * entire factor of two, and that's something we |
| 1916 | * should fix. Above that, you can misround within |
| 1917 | * the limits of your accuracy bound if you have |
| 1918 | * to. |
| 1919 | */ |
| 1920 | if (result[2] < 0x40000000) { |
| 1921 | /* Total underflow (ERANGE + UFL) is required, |
| 1922 | * and we suppress the extra bits to make |
| 1923 | * mathtest enforce that the output is really |
| 1924 | * zero. */ |
| 1925 | errstr = "underflow"; |
| 1926 | printextra = 0; |
| 1927 | } else { |
| 1928 | /* Total underflow is not required, but if the |
| 1929 | * function rounds down to zero anyway, then |
| 1930 | * we should be prepared to tolerate it. */ |
| 1931 | errstr = "?underflow"; |
| 1932 | } |
| 1933 | } else if (!(result[0] & 0x7ff00000)) { |
| 1934 | /* |
| 1935 | * If the output is denormal, we usually expect a |
| 1936 | * UFL exception, warning the user of partial |
| 1937 | * underflow. The exception is if the denormal |
| 1938 | * being returned is just one of the input values, |
| 1939 | * unchanged even in principle. I bodgily handle |
| 1940 | * this by just special-casing the functions in |
| 1941 | * question below. |
| 1942 | */ |
| 1943 | if (!strcmp(fn->name, "fmax") || |
| 1944 | !strcmp(fn->name, "fmin") || |
| 1945 | !strcmp(fn->name, "creal") || |
| 1946 | !strcmp(fn->name, "cimag")) { |
| 1947 | /* no error expected */ |
| 1948 | } else { |
| 1949 | errstr = "u"; |
| 1950 | } |
| 1951 | } else if ((result[0] & 0x7FFFFFFF) > 0x7FEFFFFF) { |
| 1952 | /* |
| 1953 | * Infinite results are usually due to overflow, |
| 1954 | * but one exception is lgamma of a negative |
| 1955 | * integer. |
| 1956 | */ |
| 1957 | if (!strcmp(fn->name, "lgamma") && |
| 1958 | (args[0] & 0x80000000) != 0 && /* negative */ |
| 1959 | is_dinteger(args)) { |
| 1960 | errstr = "ERANGE status=z"; |
| 1961 | } else { |
| 1962 | errstr = "overflow"; |
| 1963 | } |
| 1964 | printextra = 0; |
| 1965 | } |
| 1966 | } else { |
| 1967 | /* lgamma(0) is also a pole. */ |
| 1968 | if (!strcmp(fn->name, "lgamma")) { |
| 1969 | errstr = "ERANGE status=z"; |
| 1970 | printextra = 0; |
| 1971 | } |
| 1972 | } |
| 1973 | } |
| 1974 | |
| 1975 | if (!printextra || (rejected && !(rejected==1 && result[2]!=0))) { |
| 1976 | printf(" result=%08x.%08x", |
| 1977 | result[0], result[1]); |
| 1978 | } else { |
| 1979 | printf(" result=%08x.%08x.%03x", |
| 1980 | result[0], result[1], (result[2] >> 20) & 0xFFF); |
| 1981 | } |
| 1982 | if (fn->type == rred) { |
| 1983 | printf(" res2=%08x", result[3]); |
| 1984 | } |
| 1985 | break; |
| 1986 | case args1f: |
| 1987 | case args2f: |
| 1988 | case args1fcr: |
| 1989 | case rredf: |
Szabolcs Nagy | 5056a59 | 2018-04-25 09:39:45 +0100 | [diff] [blame] | 1990 | printextra = 1; |
George Lander | da55ef9 | 2015-11-19 12:05:06 +0000 | [diff] [blame] | 1991 | if (rejected == 0) { |
| 1992 | errstr = NULL; |
George Lander | da55ef9 | 2015-11-19 12:05:06 +0000 | [diff] [blame] | 1993 | if (!mpfr_zero_p(a)) { |
| 1994 | if ((result[0] & 0x7FFFFFFF) == 0) { |
| 1995 | /* |
| 1996 | * Decide whether to print the extra bits based on |
| 1997 | * just how close to zero the number is. See the |
| 1998 | * big comment in the double-precision case for |
| 1999 | * discussion. |
| 2000 | */ |
| 2001 | if (result[1] < 0x40000000) { |
| 2002 | errstr = "underflow"; |
| 2003 | printextra = 0; |
| 2004 | } else { |
| 2005 | errstr = "?underflow"; |
| 2006 | } |
| 2007 | } else if (!(result[0] & 0x7f800000)) { |
| 2008 | /* |
| 2009 | * Functions which do not report partial overflow |
| 2010 | * are listed here as special cases. (See the |
| 2011 | * corresponding double case above for a fuller |
| 2012 | * comment.) |
| 2013 | */ |
| 2014 | if (!strcmp(fn->name, "fmaxf") || |
| 2015 | !strcmp(fn->name, "fminf") || |
| 2016 | !strcmp(fn->name, "crealf") || |
| 2017 | !strcmp(fn->name, "cimagf")) { |
| 2018 | /* no error expected */ |
| 2019 | } else { |
| 2020 | errstr = "u"; |
| 2021 | } |
| 2022 | } else if ((result[0] & 0x7FFFFFFF) > 0x7F7FFFFF) { |
| 2023 | /* |
| 2024 | * Infinite results are usually due to overflow, |
| 2025 | * but one exception is lgamma of a negative |
| 2026 | * integer. |
| 2027 | */ |
| 2028 | if (!strcmp(fn->name, "lgammaf") && |
| 2029 | (args[0] & 0x80000000) != 0 && /* negative */ |
| 2030 | is_sinteger(args)) { |
| 2031 | errstr = "ERANGE status=z"; |
| 2032 | } else { |
| 2033 | errstr = "overflow"; |
| 2034 | } |
| 2035 | printextra = 0; |
| 2036 | } |
| 2037 | } else { |
| 2038 | /* lgamma(0) is also a pole. */ |
| 2039 | if (!strcmp(fn->name, "lgammaf")) { |
| 2040 | errstr = "ERANGE status=z"; |
| 2041 | printextra = 0; |
| 2042 | } |
| 2043 | } |
| 2044 | } |
| 2045 | |
| 2046 | if (!printextra || (rejected && !(rejected==1 && result[1]!=0))) { |
| 2047 | printf(" result=%08x", |
| 2048 | result[0]); |
| 2049 | } else { |
| 2050 | printf(" result=%08x.%03x", |
| 2051 | result[0], (result[1] >> 20) & 0xFFF); |
| 2052 | } |
| 2053 | if (fn->type == rredf) { |
| 2054 | printf(" res2=%08x", result[3]); |
| 2055 | } |
| 2056 | break; |
| 2057 | case semi1: /* return a double result */ |
| 2058 | case semi2: |
| 2059 | case t_ldexp: |
| 2060 | printf(" result=%08x.%08x", result[0], result[1]); |
| 2061 | break; |
| 2062 | case semi1f: |
| 2063 | case semi2f: |
| 2064 | case t_ldexpf: |
| 2065 | printf(" result=%08x", result[0]); |
| 2066 | break; |
| 2067 | case t_frexp: /* return double * int */ |
| 2068 | printf(" result=%08x.%08x res2=%08x", result[0], result[1], |
| 2069 | result[2]); |
| 2070 | break; |
| 2071 | case t_modf: /* return double * double */ |
| 2072 | printf(" result=%08x.%08x res2=%08x.%08x", |
| 2073 | result[0], result[1], result[2], result[3]); |
| 2074 | break; |
| 2075 | case t_modff: /* return float * float */ |
| 2076 | /* fall through */ |
| 2077 | case t_frexpf: /* return float * int */ |
| 2078 | printf(" result=%08x res2=%08x", result[0], result[2]); |
| 2079 | break; |
| 2080 | case classify: |
| 2081 | case classifyf: |
| 2082 | case compare: |
| 2083 | case comparef: |
| 2084 | printf(" result=%x", result[0]); |
| 2085 | break; |
| 2086 | case args1c: |
| 2087 | case args2c: |
| 2088 | if (0/* errstr */) { |
| 2089 | printf(" resultr=%08x.%08x", result[0], result[1]); |
| 2090 | printf(" resulti=%08x.%08x", result[4], result[5]); |
| 2091 | } else { |
| 2092 | printf(" resultr=%08x.%08x.%03x", |
| 2093 | result[0], result[1], (result[2] >> 20) & 0xFFF); |
| 2094 | printf(" resulti=%08x.%08x.%03x", |
| 2095 | result[4], result[5], (result[6] >> 20) & 0xFFF); |
| 2096 | } |
| 2097 | /* Underflow behaviour doesn't seem to be specified for complex arithmetic */ |
| 2098 | errstr = "?underflow"; |
| 2099 | break; |
| 2100 | case args1fc: |
| 2101 | case args2fc: |
| 2102 | if (0/* errstr */) { |
| 2103 | printf(" resultr=%08x", result[0]); |
| 2104 | printf(" resulti=%08x", result[4]); |
| 2105 | } else { |
| 2106 | printf(" resultr=%08x.%03x", |
| 2107 | result[0], (result[1] >> 20) & 0xFFF); |
| 2108 | printf(" resulti=%08x.%03x", |
| 2109 | result[4], (result[5] >> 20) & 0xFFF); |
| 2110 | } |
| 2111 | /* Underflow behaviour doesn't seem to be specified for complex arithmetic */ |
| 2112 | errstr = "?underflow"; |
| 2113 | break; |
| 2114 | } |
| 2115 | |
| 2116 | if (errstr && *(errstr+1) == '\0') { |
| 2117 | printf(" errno=0 status=%c",*errstr); |
| 2118 | } else if (errstr && *errstr == '?') { |
| 2119 | printf(" maybeerror=%s", errstr+1); |
| 2120 | } else if (errstr && errstr[0] == 'E') { |
| 2121 | printf(" errno=%s", errstr); |
| 2122 | } else { |
| 2123 | printf(" error=%s", errstr && *errstr ? errstr : "0"); |
| 2124 | } |
| 2125 | |
| 2126 | printf("\n"); |
| 2127 | |
| 2128 | vet_for_decline(fn, args, result, 0); |
| 2129 | |
| 2130 | cleanup: |
| 2131 | mpfr_clear(a); |
| 2132 | mpfr_clear(b); |
| 2133 | mpfr_clear(r); |
| 2134 | mpc_clear(ac); |
| 2135 | mpc_clear(bc); |
| 2136 | mpc_clear(rc); |
| 2137 | } |
| 2138 | |
| 2139 | void gencases(Testable *fn, int number) { |
| 2140 | int i; |
| 2141 | uint32 args[8]; |
| 2142 | |
| 2143 | float32_case(NULL); |
| 2144 | float64_case(NULL); |
| 2145 | |
| 2146 | printf("random=on\n"); /* signal to runtests.pl that the following tests are randomly generated */ |
| 2147 | for (i = 0; i < number; i++) { |
| 2148 | /* generate test point */ |
| 2149 | fn->cases(args, fn->caseparam1, fn->caseparam2); |
| 2150 | docase(fn, args); |
| 2151 | } |
| 2152 | printf("random=off\n"); |
| 2153 | } |
| 2154 | |
| 2155 | static uint32 doubletop(int x, int scale) { |
| 2156 | int e = 0x412 + scale; |
| 2157 | while (!(x & 0x100000)) |
| 2158 | x <<= 1, e--; |
| 2159 | return (e << 20) + x; |
| 2160 | } |
| 2161 | |
| 2162 | static uint32 floatval(int x, int scale) { |
| 2163 | int e = 0x95 + scale; |
| 2164 | while (!(x & 0x800000)) |
| 2165 | x <<= 1, e--; |
| 2166 | return (e << 23) + x; |
| 2167 | } |