1 /* Sum -- efficiently sum a list of floating-point numbers 2 3 Copyright 2014-2023 Free Software Foundation, Inc. 4 Contributed by the AriC and Caramba projects, INRIA. 5 6 This file is part of the GNU MPFR Library. 7 8 The GNU MPFR Library is free software; you can redistribute it and/or modify 9 it under the terms of the GNU Lesser General Public License as published by 10 the Free Software Foundation; either version 3 of the License, or (at your 11 option) any later version. 12 13 The GNU MPFR Library is distributed in the hope that it will be useful, but 14 WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY 15 or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public 16 License for more details. 17 18 You should have received a copy of the GNU Lesser General Public License 19 along with the GNU MPFR Library; see the file COPYING.LESSER. If not, see 20 https://www.gnu.org/licenses/ or write to the Free Software Foundation, Inc., 21 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA. */ 22 23 #define MPFR_NEED_LONGLONG_H 24 #include "mpfr-impl.h" 25 26 /* Note: In the prototypes, one uses 27 * 28 * const mpfr_ptr *x i.e.: __mpfr_struct *const *x 29 * 30 * instead of 31 * 32 * const mpfr_srcptr *x i.e.: const __mpfr_struct *const *x 33 * 34 * because here one has a double indirection and the type matching rules 35 * from the C standard in such a case are stricter and they would yield 36 * annoying errors for the user in practice. See: 37 * 38 * Why can't I pass a char ** to a function which expects a const char **? 39 * 40 * in the comp.lang.c FAQ: 41 * 42 * https://c-faq.com/ansi/constmismatch.html 43 */ 44 45 /* See the doc/sum.txt file for the algorithm and a part of its proof 46 (this will later go into algorithms.tex). 47 48 TODO [VL, after a discussion with James Demmel]: Compared to 49 James Demmel and Yozo Hida, Fast and accurate floating-point summation 50 with application to computational geometry, Numerical Algorithms, 51 volume 37, number 1-4, pages 101--112, 2004. 52 sorting is not necessary here. It is not done because in the most common 53 cases (where big cancellations are rare), it would take time and be 54 useless. However, the lack of sorting increases the worst case complexity. 55 For instance, consider many inputs that cancel one another (two by two). 56 One would need n/2 iterations, where each iteration reads the exponent 57 of each input, therefore n*n/2 read operations. Using a worst-case sort 58 in O(n log n) could give a O(n log n) worst-case complexity. As we don't 59 want to slow down the most common cases, this could be done at the 3rd 60 iteration. But are there practical applications which would be used as 61 tests? 62 63 Note: see the following paper and its references: 64 http://www.acsel-lab.com/arithmetic/arith21/papers/p54.pdf 65 (J. Demmel and H. D. Nguyen, Fast Reproducible Floating-Point Summation) 66 VL: This is very different: 67 In MPFR In the paper & references 68 arbitrary precision fixed precision 69 correct rounding just reproducible rounding 70 integer operations floating-point operations 71 sequential parallel (& sequential) 72 */ 73 74 #ifdef MPFR_COV_CHECK 75 int __gmpfr_cov_sum_tmd[MPFR_RND_MAX][2][2][3][2][2] = { 0 }; 76 #endif 77 78 /* Update minexp (V) after detecting a potential integer overflow in 79 extreme cases (only a 32-bit ABI may be concerned in practice). 80 Instead of an assertion failure below, we could 81 1. check that the ulp of each regular input has an exponent >= MPFR_EXP_MIN 82 (with an assertion failure if this is not the case); 83 2. set minexp to MPFR_EXP_MIN and shift the accumulator accordingly 84 (the sum will then be exact). 85 However, such cases, which involve huge precisions, will probably 86 never occur in practice (at least with a 64-bit ABI) and are not 87 easily testable due to these huge precisions. Moreover, switching 88 to a 64-bit ABI would be a better solution for such computations. 89 So, let's leave this unimplemented. */ 90 #define SAFE_SUB(V,E,SH) \ 91 do \ 92 { \ 93 mpfr_prec_t sh = (SH); \ 94 MPFR_ASSERTN ((E) >= MPFR_EXP_MIN + sh); \ 95 V = (E) - sh; \ 96 } \ 97 while (0) 98 99 /* Function sum_raw 100 * ================ 101 * 102 * Accumulate a new [minexp,maxexp[ block into (wp,ws). If e and err denote 103 * the exponents of the computed result and of the error bound respectively, 104 * while e - err is less than some given bound (due to cancellation), shift 105 * the accumulator and reiterate. 106 * 107 * Inputs: 108 * wp: pointer to the accumulator (least significant limb first). 109 * ws: size of the accumulator (in limbs). 110 * wq: precision of the accumulator (ws * GMP_NUMB_BITS). 111 * x: array of the input numbers. 112 * n: size of this array (number of inputs, regular or not). 113 * minexp: exponent of the least significant bit of the first block. 114 * maxexp: exponent of the first block (exponent of its MSB + 1). 115 * tp: pointer to a temporary area (pre-allocated). 116 * ts: size of this temporary area. 117 * logn: ceil(log2(rn)), where rn is the number of regular inputs. 118 * prec: lower bound for e - err (as described above). 119 * ep: pointer to mpfr_exp_t (see below), or a null pointer. 120 * minexpp: pointer to mpfr_exp_t (see below), or a null pointer. 121 * maxexpp: pointer to mpfr_exp_t (see below), or a null pointer. 122 * 123 * Preconditions: 124 * prec >= 1 125 * wq >= logn + prec + 2 126 * 127 * This function returns 0 if the accumulator is 0 (which implies that 128 * the exact sum for this sum_raw invocation is 0), otherwise the number 129 * of cancelled bits (>= 1), defined as the number of identical bits on 130 * the most significant part of the accumulator. In the latter case, it 131 * also returns the following data in variables passed by reference, if 132 * the pointers are not NULL: 133 * - in ep: the exponent e of the computed result; 134 * - in minexpp: the last value of minexp; 135 * - in maxexpp: the new value of maxexp (for the next iteration after 136 * the first invocation of sum_raw in the main code). 137 * 138 * Notes: 139 * - minexp is also the exponent of the least significant bit of the 140 * accumulator; 141 * - the temporary area must be large enough to hold a shifted input 142 * block, and the value of ts is used only when the full assertions 143 * are checked (i.e. with the --enable-assert configure option), to 144 * check that a buffer overflow doesn't occur; 145 * - contrary to the returned value of minexp (the value in the last 146 * iteration), the returned value of maxexp is the one for the next 147 * iteration (= maxexp2 of the last iteration). 148 */ 149 static mpfr_prec_t 150 sum_raw (mp_limb_t *wp, mp_size_t ws, mpfr_prec_t wq, const mpfr_ptr *x, 151 unsigned long n, mpfr_exp_t minexp, mpfr_exp_t maxexp, 152 mp_limb_t *tp, mp_size_t ts, int logn, mpfr_prec_t prec, 153 mpfr_exp_t *ep, mpfr_exp_t *minexpp, mpfr_exp_t *maxexpp) 154 { 155 MPFR_LOG_FUNC 156 (("ws=%Pd ts=%Pd prec=%Pd", (mpfr_prec_t) ws, (mpfr_prec_t) ts, prec), 157 ("", 0)); 158 159 /* The C code below requires prec >= 0 due to the use of unsigned 160 integer arithmetic on it. Actually the computation makes sense 161 only with prec >= 1 (otherwise one can't even know the sign of 162 the result), hence the following assertion. */ 163 MPFR_ASSERTD (prec >= 1); 164 165 /* Consistency check. */ 166 MPFR_ASSERTD (wq == (mpfr_prec_t) ws * GMP_NUMB_BITS); 167 168 /* The following precondition together with prec >= 1 will imply: 169 minexp - shiftq < maxexp2, as required by the algorithm. */ 170 MPFR_ASSERTD (wq >= logn + prec + 2); 171 172 while (1) 173 { 174 mpfr_exp_t maxexp2 = MPFR_EXP_MIN; 175 unsigned long i; 176 177 MPFR_LOG_MSG (("sum_raw loop: " 178 "maxexp=%" MPFR_EXP_FSPEC "d " 179 "minexp=%" MPFR_EXP_FSPEC "d\n", 180 (mpfr_eexp_t) maxexp, (mpfr_eexp_t) minexp)); 181 182 MPFR_ASSERTD (maxexp > minexp); 183 184 for (i = 0; i < n; i++) 185 if (! MPFR_IS_SINGULAR (x[i])) /* Step 1 (see sum_raw in sum.txt) */ 186 { 187 mp_limb_t *dp, *vp; 188 mp_size_t ds, vs, vds; 189 mpfr_exp_t xe, vd; 190 mpfr_prec_t xq; 191 int tr; 192 193 xe = MPFR_GET_EXP (x[i]); 194 xq = MPFR_GET_PREC (x[i]); 195 196 vp = MPFR_MANT (x[i]); 197 vs = MPFR_PREC2LIMBS (xq); 198 vd = xe - vs * GMP_NUMB_BITS - minexp; 199 /* vd is the exponent of the least significant represented bit of 200 x[i] (including the trailing bits, whose value is 0) minus the 201 exponent of the least significant bit of the accumulator. To 202 make the code simpler, we won't try to filter out the trailing 203 bits of x[i]. */ 204 205 /* Steps 2, 3, 4 (see sum_raw in sum.txt) */ 206 207 if (vd < 0) 208 { 209 /* This covers the following cases: 210 * [-+- accumulator ---] 211 * [---|----- x[i] ------|--] 212 * | [----- x[i] --|--] 213 * | |[----- x[i] -----] 214 * | | [----- x[i] -----] 215 * maxexp minexp 216 */ 217 218 /* Step 2 for subcase vd < 0 */ 219 220 if (xe <= minexp) 221 { 222 /* x[i] is entirely after the LSB of the accumulator, 223 so that it will be ignored at this iteration. */ 224 if (xe > maxexp2) 225 { 226 maxexp2 = xe; 227 /* And since the exponent of x[i] is valid... */ 228 MPFR_ASSERTD (maxexp2 >= MPFR_EMIN_MIN); 229 } 230 continue; 231 } 232 233 /* Step 3 for subcase vd < 0 */ 234 235 /* If some significant bits of x[i] are after the LSB of the 236 accumulator, then maxexp2 will necessarily be minexp. */ 237 if (MPFR_LIKELY (xe - xq < minexp)) 238 maxexp2 = minexp; 239 240 /* Step 4 for subcase vd < 0 */ 241 242 /* We need to ignore the least |vd| significant bits of x[i]. 243 First, let's ignore the least vds = |vd| / GMP_NUMB_BITS 244 limbs. */ 245 vd = - vd; 246 vds = vd / GMP_NUMB_BITS; 247 vs -= vds; 248 MPFR_ASSERTD (vs > 0); /* see xe <= minexp test above */ 249 vp += vds; 250 vd -= vds * GMP_NUMB_BITS; 251 MPFR_ASSERTD (vd >= 0 && vd < GMP_NUMB_BITS); 252 253 if (xe > maxexp) 254 { 255 vs -= (xe - maxexp) / GMP_NUMB_BITS; 256 MPFR_ASSERTD (vs > 0); 257 tr = (xe - maxexp) % GMP_NUMB_BITS; 258 } 259 else 260 tr = 0; 261 262 if (vd != 0) 263 { 264 MPFR_ASSERTD (vs <= ts); 265 mpn_rshift (tp, vp, vs, vd); 266 vp = tp; 267 tr += vd; 268 if (tr >= GMP_NUMB_BITS) 269 { 270 vs--; 271 tr -= GMP_NUMB_BITS; 272 } 273 MPFR_ASSERTD (vs >= 1); 274 MPFR_ASSERTD (tr >= 0 && tr < GMP_NUMB_BITS); 275 if (tr != 0) 276 { 277 tp[vs-1] &= MPFR_LIMB_MASK (GMP_NUMB_BITS - tr); 278 tr = 0; 279 } 280 /* Truncation has now been taken into account. */ 281 MPFR_ASSERTD (tr == 0); 282 } 283 284 dp = wp; 285 ds = ws; 286 } 287 else /* vd >= 0 */ 288 { 289 /* This covers the following cases: 290 * [-+- accumulator ---] 291 * [- x[i] -] | | 292 * [---|-- x[i] ------] | 293 * [------|-- x[i] ---------] 294 * | [- x[i] -] | 295 * maxexp minexp 296 */ 297 298 /* Steps 2 and 3 for subcase vd >= 0 */ 299 300 MPFR_ASSERTD (xe - xq >= minexp); /* see definition of vd */ 301 302 /* Step 4 for subcase vd >= 0 */ 303 304 /* We need to ignore the least vd significant bits 305 of the accumulator. First, let's ignore the least 306 vds = vd / GMP_NUMB_BITS limbs. -> (dp,ds) */ 307 vds = vd / GMP_NUMB_BITS; 308 ds = ws - vds; 309 if (ds <= 0) 310 continue; 311 dp = wp + vds; 312 vd -= vds * GMP_NUMB_BITS; 313 MPFR_ASSERTD (vd >= 0 && vd < GMP_NUMB_BITS); 314 315 /* The low part of x[i] (to be determined) will have to be 316 shifted vd bits to the left if vd != 0. */ 317 318 if (xe > maxexp) 319 { 320 vs -= (xe - maxexp) / GMP_NUMB_BITS; 321 if (vs <= 0) 322 continue; 323 tr = (xe - maxexp) % GMP_NUMB_BITS; 324 } 325 else 326 tr = 0; 327 328 MPFR_ASSERTD (tr >= 0 && tr < GMP_NUMB_BITS && vs > 0); 329 330 /* We need to consider the least significant vs limbs of x[i] 331 except the most significant tr bits. */ 332 333 if (vd != 0) 334 { 335 mp_limb_t carry; 336 337 MPFR_ASSERTD (vs <= ts); 338 carry = mpn_lshift (tp, vp, vs, vd); 339 tr -= vd; 340 if (tr < 0) 341 { 342 tr += GMP_NUMB_BITS; 343 MPFR_ASSERTD (vs + 1 <= ts); 344 tp[vs++] = carry; 345 } 346 MPFR_ASSERTD (tr >= 0 && tr < GMP_NUMB_BITS); 347 vp = tp; 348 } 349 } /* vd >= 0 */ 350 351 MPFR_ASSERTD (vs > 0 && vs <= ds); 352 353 /* We can't truncate the most significant limb of the input 354 (in case it hasn't been shifted to the temporary area). 355 So, let's ignore it now. It will be taken into account 356 via carry propagation after the addition. */ 357 if (tr != 0) 358 vs--; 359 360 /* Step 5 (see sum_raw in sum.txt) */ 361 362 if (MPFR_IS_POS (x[i])) 363 { 364 mp_limb_t carry; 365 366 carry = vs > 0 ? mpn_add_n (dp, dp, vp, vs) : 0; 367 MPFR_ASSERTD (carry <= 1); 368 if (tr != 0) 369 carry += vp[vs] & MPFR_LIMB_MASK (GMP_NUMB_BITS - tr); 370 if (ds > vs) 371 mpn_add_1 (dp + vs, dp + vs, ds - vs, carry); 372 } 373 else 374 { 375 mp_limb_t borrow; 376 377 borrow = vs > 0 ? mpn_sub_n (dp, dp, vp, vs) : 0; 378 MPFR_ASSERTD (borrow <= 1); 379 if (tr != 0) 380 borrow += vp[vs] & MPFR_LIMB_MASK (GMP_NUMB_BITS - tr); 381 if (ds > vs) 382 mpn_sub_1 (dp + vs, dp + vs, ds - vs, borrow); 383 } 384 } 385 386 { 387 mpfr_prec_t cancel; /* number of cancelled bits */ 388 mp_size_t wi; /* index in the accumulator */ 389 mp_limb_t a, b; 390 int cnt; 391 392 cancel = 0; 393 wi = ws - 1; 394 MPFR_ASSERTD (wi >= 0); 395 a = wp[wi] >> (GMP_NUMB_BITS - 1) ? MPFR_LIMB_MAX : MPFR_LIMB_ZERO; 396 397 while (wi >= 0) 398 if ((b = wp[wi]) == a) 399 { 400 cancel += GMP_NUMB_BITS; 401 wi--; 402 } 403 else 404 { 405 b ^= a; 406 MPFR_ASSERTD (b != 0); 407 count_leading_zeros (cnt, b); 408 cancel += cnt; 409 break; 410 } 411 412 if (wi >= 0 || a != MPFR_LIMB_ZERO) /* accumulator != 0 */ 413 { 414 mpfr_exp_t e; /* exponent of the computed result */ 415 mpfr_exp_t err; /* exponent of the error bound */ 416 417 MPFR_LOG_MSG (("accumulator %s 0, cancel=%Pd\n", 418 a != MPFR_LIMB_ZERO ? "<" : ">", cancel)); 419 420 MPFR_ASSERTD (cancel > 0); 421 e = minexp + wq - cancel; 422 MPFR_ASSERTD (e >= minexp); 423 err = maxexp2 + logn; /* OK even if maxexp2 == MPFR_EXP_MIN */ 424 425 /* The absolute value of the truncated sum is in the binade 426 [2^(e-1),2^e] (closed on both ends due to two's complement). 427 The error is strictly less than 2^err (and is 0 if 428 maxexp2 == MPFR_EXP_MIN). */ 429 430 /* This basically tests whether err <= e - prec without 431 potential integer overflow (since prec >= 0)... 432 Note that the maxexp2 == MPFR_EXP_MIN test is there just for 433 the potential corner case e - prec < MPFR_EXP_MIN + logn. 434 Such corner cases, involving specific huge-precision numbers, 435 are probably not supported in many places in MPFR, but this 436 test doesn't hurt... */ 437 if (maxexp2 == MPFR_EXP_MIN || 438 (err <= e && SAFE_DIFF (mpfr_uexp_t, e, err) >= prec)) 439 { 440 MPFR_LOG_MSG (("(err=%" MPFR_EXP_FSPEC "d) <= (e=%" 441 MPFR_EXP_FSPEC "d) - (prec=%Pd)\n", 442 (mpfr_eexp_t) err, (mpfr_eexp_t) e, prec)); 443 /* To avoid tests or copies, we consider the only two cases 444 that will occur in sum_aux. */ 445 MPFR_ASSERTD ((ep != NULL && 446 minexpp != NULL && 447 maxexpp != NULL) || 448 (ep == NULL && 449 minexpp == NULL && 450 maxexpp == NULL)); 451 if (ep != NULL) 452 { 453 *ep = e; 454 *minexpp = minexp; 455 *maxexpp = maxexp2; 456 } 457 MPFR_LOG_MSG (("return with minexp=%" MPFR_EXP_FSPEC 458 "d maxexp2=%" MPFR_EXP_FSPEC "d%s\n", 459 (mpfr_eexp_t) minexp, (mpfr_eexp_t) maxexp2, 460 maxexp2 == MPFR_EXP_MIN ? 461 " (MPFR_EXP_MIN)" : "")); 462 return cancel; 463 } 464 else 465 { 466 mpfr_exp_t diffexp; 467 mpfr_prec_t shiftq; 468 mpfr_size_t shifts; 469 int shiftc; 470 471 MPFR_LOG_MSG (("e=%" MPFR_EXP_FSPEC "d err=%" MPFR_EXP_FSPEC 472 "d maxexp2=%" MPFR_EXP_FSPEC "d%s\n", 473 (mpfr_eexp_t) e, (mpfr_eexp_t) err, 474 (mpfr_eexp_t) maxexp2, 475 maxexp2 == MPFR_EXP_MIN ? 476 " (MPFR_EXP_MIN)" : "")); 477 478 diffexp = err - e; 479 if (diffexp < 0) 480 diffexp = 0; 481 /* diffexp = max(0, err - e) */ 482 483 MPFR_LOG_MSG (("diffexp=%" MPFR_EXP_FSPEC "d\n", 484 (mpfr_eexp_t) diffexp)); 485 486 MPFR_ASSERTD (diffexp < cancel - 2); 487 shiftq = cancel - 2 - (mpfr_prec_t) diffexp; 488 /* equivalent to: minexp + wq - 2 - max(e,err) */ 489 MPFR_ASSERTD (shiftq > 0); 490 shifts = shiftq / GMP_NUMB_BITS; 491 shiftc = shiftq % GMP_NUMB_BITS; 492 MPFR_LOG_MSG (("shiftq = %Pd = %Pd * GMP_NUMB_BITS + %d\n", 493 shiftq, (mpfr_prec_t) shifts, shiftc)); 494 if (MPFR_LIKELY (shiftc != 0)) 495 mpn_lshift (wp + shifts, wp, ws - shifts, shiftc); 496 else 497 mpn_copyd (wp + shifts, wp, ws - shifts); 498 MPN_ZERO (wp, shifts); 499 /* Compute minexp = minexp - shiftq safely. */ 500 SAFE_SUB (minexp, minexp, shiftq); 501 MPFR_ASSERTD (minexp < maxexp2); 502 } 503 } 504 else if (maxexp2 == MPFR_EXP_MIN) 505 { 506 MPFR_LOG_MSG (("accumulator = 0, maxexp2 = MPFR_EXP_MIN\n", 0)); 507 return 0; 508 } 509 else 510 { 511 MPFR_LOG_MSG (("accumulator = 0, reiterate\n", 0)); 512 /* Compute minexp = maxexp2 - (wq - (logn + 1)) safely. */ 513 SAFE_SUB (minexp, maxexp2, wq - (logn + 1)); 514 /* Note: the logn + 1 corresponds to cq in the main code. */ 515 } 516 } 517 518 maxexp = maxexp2; 519 } 520 } 521 522 /**********************************************************************/ 523 524 /* Generic case: all the inputs are finite numbers, 525 with at least 3 regular numbers. */ 526 static int 527 sum_aux (mpfr_ptr sum, const mpfr_ptr *x, unsigned long n, mpfr_rnd_t rnd, 528 mpfr_exp_t maxexp, unsigned long rn) 529 { 530 mp_limb_t *sump; 531 mp_limb_t *tp; /* pointer to a temporary area */ 532 mp_limb_t *wp; /* pointer to the accumulator */ 533 mp_size_t ts; /* size of the temporary area, in limbs */ 534 mp_size_t ws; /* size of the accumulator, in limbs */ 535 mp_size_t zs; /* size of the TMD accumulator, in limbs */ 536 mpfr_prec_t wq; /* size of the accumulator, in bits */ 537 int logn; /* ceil(log2(rn)) */ 538 int cq; 539 mpfr_prec_t sq; 540 int inex; 541 MPFR_TMP_DECL (marker); 542 543 MPFR_LOG_FUNC 544 (("n=%lu rnd=%d maxexp=%" MPFR_EXP_FSPEC "d rn=%lu", 545 n, rnd, (mpfr_eexp_t) maxexp, rn), 546 ("sum[%Pd]=%.*Rg", mpfr_get_prec (sum), mpfr_log_prec, sum)); 547 548 MPFR_ASSERTD (rn >= 3 && rn <= n); 549 550 /* In practice, no integer overflow on the exponent. */ 551 MPFR_STAT_STATIC_ASSERT (MPFR_EXP_MAX - MPFR_EMAX_MAX >= 552 sizeof (unsigned long) * CHAR_BIT); 553 554 /* Set up some variables and the accumulator. */ 555 556 sump = MPFR_MANT (sum); 557 558 /* rn is the number of regular inputs (the singular ones will be 559 ignored). Compute logn = ceil(log2(rn)). */ 560 logn = MPFR_INT_CEIL_LOG2 (rn); 561 MPFR_ASSERTD (logn >= 2); 562 563 MPFR_LOG_MSG (("logn=%d maxexp=%" MPFR_EXP_FSPEC "d\n", 564 logn, (mpfr_eexp_t) maxexp)); 565 566 sq = MPFR_GET_PREC (sum); 567 cq = logn + 1; 568 569 /* First determine the size of the accumulator. 570 * cq + sq + logn + 2 >= logn + sq + 5, which will be used later. 571 * The assertion wq - cq - sq >= 4 is another way to check that. 572 */ 573 ws = MPFR_PREC2LIMBS (cq + sq + logn + 2); 574 wq = (mpfr_prec_t) ws * GMP_NUMB_BITS; 575 MPFR_ASSERTD (wq - cq - sq >= 4); 576 577 /* TODO: timings, comparing with a larger zs. */ 578 zs = MPFR_PREC2LIMBS (wq - sq); 579 580 MPFR_LOG_MSG (("cq=%d sq=%Pd logn=%d wq=%Pd\n", cq, sq, logn, wq)); 581 582 /* An input block will have up to wq - cq bits, and its shifted value 583 (to be correctly aligned) may take GMP_NUMB_BITS - 1 additional bits. */ 584 ts = MPFR_PREC2LIMBS (wq - cq + GMP_NUMB_BITS - 1); 585 586 MPFR_TMP_MARK (marker); 587 588 /* Note: If the TMD does not occur, which should be the case for most 589 sums, allocating zs limbs is not necessary. However, we choose to 590 do this now (thus in all cases) because zs is very small, so that 591 the difference on the memory footprint will not be noticeable. 592 More precisely, zs is at most 2 in practice with the current code; 593 we may want to increase it in order to avoid performance issues in 594 some unlikely corner cases, but even in this case, it will remain 595 small. 596 One will have: 597 [------ ts ------][------ ws ------][- zs -] 598 The following would probably be better: 599 [------ ts ------] [------ ws ------] 600 [- zs -] 601 i.e. where the TMD accumulator (partially or completely) takes 602 some unneeded part of the temporary area in order to improve 603 data locality. But 604 * in low precision, data locality is regarded as ensured even 605 with the actual choice; 606 * in high precision, data locality for TMD resolution may not 607 be that important. 608 */ 609 tp = MPFR_TMP_LIMBS_ALLOC (ts + ws + zs); 610 wp = tp + ts; 611 612 MPN_ZERO (wp, ws); /* zero the accumulator */ 613 614 { 615 mpfr_exp_t minexp; /* exponent of the LSB of the block for sum_raw */ 616 mpfr_prec_t cancel; /* number of cancelled bits */ 617 mpfr_exp_t e; /* temporary exponent of the result */ 618 mpfr_exp_t u; /* temporary exponent of the ulp (quantum) */ 619 mp_limb_t lbit; /* last bit (useful if even rounding) */ 620 mp_limb_t rbit; /* rounding bit (corrected in halfway case) */ 621 int corr; /* correction term (from -1 to 2) */ 622 int sd, sh; /* shift counts */ 623 mp_size_t sn; /* size of the output number */ 624 int tmd; /* 0: the TMD does not occur 625 1: the TMD occurs on a machine number 626 2: the TMD occurs on a midpoint */ 627 int neg; /* 0 if positive sum, 1 if negative */ 628 int sgn; /* +1 if positive sum, -1 if negative */ 629 630 MPFR_LOG_MSG (("Compute an approximation with sum_raw...\n", 0)); 631 632 /* Compute minexp = maxexp - (wq - cq) safely. */ 633 SAFE_SUB (minexp, maxexp, wq - cq); 634 MPFR_ASSERTD (wq >= logn + sq + 5); 635 cancel = sum_raw (wp, ws, wq, x, n, minexp, maxexp, tp, ts, 636 logn, sq + 3, &e, &minexp, &maxexp); 637 638 if (MPFR_UNLIKELY (cancel == 0)) 639 { 640 /* The exact sum is zero. Since not all inputs are 0, the sum 641 * is +0 except in MPFR_RNDD, as specified according to the 642 * IEEE 754 rules for the addition of two numbers. 643 */ 644 MPFR_SET_SIGN (sum, (rnd != MPFR_RNDD ? 645 MPFR_SIGN_POS : MPFR_SIGN_NEG)); 646 MPFR_SET_ZERO (sum); 647 MPFR_TMP_FREE (marker); 648 MPFR_RET (0); 649 } 650 651 /* The absolute value of the truncated sum is in the binade 652 [2^(e-1),2^e] (closed on both ends due to two's complement). 653 The error is strictly less than 2^(maxexp + logn) (and is 0 654 if maxexp == MPFR_EXP_MIN). */ 655 656 u = e - sq; /* e being the exponent, u is the ulp of the target */ 657 658 /* neg = 1 if negative, 0 if positive. */ 659 neg = wp[ws-1] >> (GMP_NUMB_BITS - 1); 660 MPFR_ASSERTD (neg == 0 || neg == 1); 661 662 sgn = neg ? -1 : 1; 663 MPFR_ASSERTN (sgn == (neg ? MPFR_SIGN_NEG : MPFR_SIGN_POS)); 664 665 MPFR_LOG_MSG (("neg=%d sgn=%d cancel=%Pd" 666 " e=%" MPFR_EXP_FSPEC "d" 667 " u=%" MPFR_EXP_FSPEC "d" 668 " maxexp=%" MPFR_EXP_FSPEC "d%s\n", 669 neg, sgn, cancel, (mpfr_eexp_t) e, (mpfr_eexp_t) u, 670 (mpfr_eexp_t) maxexp, 671 maxexp == MPFR_EXP_MIN ? " (MPFR_EXP_MIN)" : "")); 672 673 if (rnd == MPFR_RNDF) 674 { 675 /* Rounding the approximate value to nearest (ties don't matter) is 676 sufficient. We need to get the rounding bit; the code is similar 677 to a part from the generic code (here, corr = rbit). */ 678 if (MPFR_LIKELY (u > minexp)) 679 { 680 mpfr_prec_t tq; 681 mp_size_t wi; 682 int td; 683 684 tq = u - minexp; 685 MPFR_ASSERTD (tq > 0); /* number of trailing bits */ 686 MPFR_LOG_MSG (("tq=%Pd\n", tq)); 687 688 wi = tq / GMP_NUMB_BITS; 689 td = tq % GMP_NUMB_BITS; 690 corr = td >= 1 ? ((wp[wi] >> (td - 1)) & MPFR_LIMB_ONE) : 691 (MPFR_ASSERTD (wi >= 1), wp[wi-1] >> (GMP_NUMB_BITS - 1)); 692 } 693 else 694 corr = 0; 695 inex = 0; /* not meaningful, but needs to have a value */ 696 } 697 else /* rnd != MPFR_RNDF */ 698 { 699 if (MPFR_LIKELY (u > minexp)) 700 { 701 mpfr_prec_t tq; 702 mp_size_t wi; 703 int td; 704 705 tq = u - minexp; 706 MPFR_ASSERTD (tq > 0); /* number of trailing bits */ 707 MPFR_LOG_MSG (("tq=%Pd\n", tq)); 708 709 wi = tq / GMP_NUMB_BITS; 710 711 /* Determine the rounding bit, which is represented. */ 712 td = tq % GMP_NUMB_BITS; 713 lbit = (wp[wi] >> td) & MPFR_LIMB_ONE; 714 rbit = td >= 1 ? ((wp[wi] >> (td - 1)) & MPFR_LIMB_ONE) : 715 (MPFR_ASSERTD (wi >= 1), wp[wi-1] >> (GMP_NUMB_BITS - 1)); 716 MPFR_ASSERTD (rbit == 0 || rbit == 1); 717 MPFR_LOG_MSG (("rbit=%d\n", (int) rbit)); 718 719 if (maxexp == MPFR_EXP_MIN) 720 { 721 /* The sum in the accumulator is exact. Determine inex: 722 inex = 0 if the final sum is exact, else 1, i.e. 723 inex = rounding bit || sticky bit. In round to nearest, 724 also determine the rounding direction: obtained from 725 the rounding bit possibly except in halfway cases. 726 Halfway cases are rounded toward -inf iff the last bit 727 of the truncated significand in two's complement is 0 728 (in precision > 1, because the parity after rounding is 729 the same in two's complement and sign + magnitude; in 730 precision 1, one checks that the rule works for both 731 positive (lbit == 1) and negative (lbit == 0) numbers, 732 rounding halfway cases away from zero). */ 733 if (MPFR_LIKELY (rbit == 0 || (rnd == MPFR_RNDN && lbit == 0))) 734 { 735 /* We need to determine the sticky bit, either to set inex 736 (if the rounding bit is 0) or to possibly "correct" rbit 737 (round to nearest, halfway case rounded downward) from 738 which the rounding direction will be determined. */ 739 MPFR_LOG_MSG (("Determine the sticky bit...\n", 0)); 740 741 inex = td >= 2 ? (wp[wi] & MPFR_LIMB_MASK (td - 1)) != 0 742 : td == 0 ? 743 (MPFR_ASSERTD (wi >= 1), 744 (wp[--wi] & MPFR_LIMB_MASK (GMP_NUMB_BITS - 1)) != 0) 745 : 0; 746 747 if (!inex) 748 { 749 while (!inex && wi > 0) 750 inex = wp[--wi] != 0; 751 if (!inex && rbit != 0) 752 { 753 /* sticky bit = 0, rounding bit = 1, 754 i.e. halfway case, which will be 755 rounded downward (see earlier if). */ 756 MPFR_ASSERTD (rnd == MPFR_RNDN); 757 inex = 1; 758 rbit = 0; /* even rounding downward */ 759 MPFR_LOG_MSG (("Halfway case rounded downward;" 760 " set inex=1 rbit=0\n", 0)); 761 } 762 } 763 } 764 else 765 inex = 1; 766 tmd = 0; /* We can round correctly -> no TMD. */ 767 } 768 else /* maxexp > MPFR_EXP_MIN */ 769 { 770 mpfr_exp_t d; 771 mp_limb_t limb, mask; 772 int nbits; 773 774 /* Since maxexp was set to either the exponent of a x[i] or 775 to minexp... */ 776 MPFR_ASSERTD (maxexp >= MPFR_EMIN_MIN || maxexp == minexp); 777 778 inex = 1; /* We do not know whether the sum is exact. */ 779 780 MPFR_ASSERTD (u <= MPFR_EMAX_MAX && u <= minexp + wq); 781 d = u - (maxexp + logn); /* representable */ 782 MPFR_ASSERTD (d >= 3); /* due to prec = sq + 3 in sum_raw */ 783 784 /* Let's see whether the TMD occurs by looking at the d bits 785 following the ulp bit, or the d-1 bits after the rounding 786 bit. */ 787 788 /* First chunk after the rounding bit... It starts at: 789 (wi,td-2) if td >= 2, 790 (wi-1,td-2+GMP_NUMB_BITS) if td < 2. */ 791 if (td == 0) 792 { 793 MPFR_ASSERTD (wi >= 1); 794 limb = wp[--wi]; 795 mask = MPFR_LIMB_MASK (GMP_NUMB_BITS - 1); 796 nbits = GMP_NUMB_BITS; 797 } 798 else if (td == 1) 799 { 800 limb = wi >= 1 ? wp[--wi] : MPFR_LIMB_ZERO; 801 mask = MPFR_LIMB_MAX; 802 nbits = GMP_NUMB_BITS + 1; 803 } 804 else /* td >= 2 */ 805 { 806 MPFR_ASSERTD (td >= 2); 807 limb = wp[wi]; 808 mask = MPFR_LIMB_MASK (td - 1); 809 nbits = td; 810 } 811 812 /* nbits: number of bits of the first chunk + 1 813 (the +1 is for the rounding bit). */ 814 815 if (nbits > d) 816 { 817 /* Some low significant bits must be ignored. */ 818 limb >>= nbits - d; 819 mask >>= nbits - d; 820 d = 0; 821 } 822 else 823 { 824 d -= nbits; 825 MPFR_ASSERTD (d >= 0); 826 } 827 828 limb &= mask; 829 tmd = 830 limb == MPFR_LIMB_ZERO ? 831 (rbit == 0 ? 1 : rnd == MPFR_RNDN ? 2 : 0) : 832 limb == mask ? 833 (limb = MPFR_LIMB_MAX, 834 rbit != 0 ? 1 : rnd == MPFR_RNDN ? 2 : 0) : 0; 835 836 while (tmd != 0 && d != 0) 837 { 838 mp_limb_t limb2; 839 840 MPFR_ASSERTD (d > 0); 841 if (wi == 0) 842 { 843 /* The non-represented bits are 0's. */ 844 if (limb != MPFR_LIMB_ZERO) 845 tmd = 0; 846 break; 847 } 848 MPFR_ASSERTD (wi > 0); 849 limb2 = wp[--wi]; 850 if (d < GMP_NUMB_BITS) 851 { 852 int c = GMP_NUMB_BITS - d; 853 MPFR_ASSERTD (c > 0 && c < GMP_NUMB_BITS); 854 if ((limb2 >> c) != (limb >> c)) 855 tmd = 0; 856 break; 857 } 858 if (limb2 != limb) 859 tmd = 0; 860 d -= GMP_NUMB_BITS; 861 } 862 } 863 } 864 else /* u <= minexp */ 865 { 866 /* The exact value of the accumulator will be copied. 867 * The TMD occurs if and only if there are bits still 868 * not taken into account, and if it occurs, this is 869 * necessarily on a machine number (-> tmd = 1). 870 */ 871 lbit = u == minexp ? wp[0] & MPFR_LIMB_ONE : 0; 872 rbit = 0; 873 inex = tmd = maxexp != MPFR_EXP_MIN; 874 } 875 876 MPFR_ASSERTD (rbit == 0 || rbit == 1); 877 878 MPFR_LOG_MSG (("tmd=%d lbit=%d rbit=%d inex=%d neg=%d\n", 879 tmd, (int) lbit, (int) rbit, inex, neg)); 880 881 /* Here, if the final sum is known to be exact, inex = 0, otherwise 882 * inex = 1. We have a truncated significand, a trailing term t such 883 * that 0 <= t < 1 ulp, and an error on the trailing term bounded by 884 * t' in absolute value. Thus the error e on the truncated significand 885 * satisfies -t' <= e < 1 ulp + t'. Thus one has 4 correction cases 886 * denoted by a corr value between -1 and 2 depending on e, neg, rbit, 887 * and the rounding mode: 888 * -1: equivalent to nextbelow; 889 * 0: the truncated significand is not corrected; 890 * 1: add 1 ulp; 891 * 2: add 1 ulp, then nextabove. 892 * The nextbelow and nextabove are used here since there may be a 893 * change of the binade. 894 */ 895 896 if (tmd == 0) /* no TMD */ 897 { 898 switch (rnd) 899 { 900 case MPFR_RNDD: 901 corr = 0; 902 break; 903 case MPFR_RNDU: 904 corr = inex; 905 break; 906 case MPFR_RNDZ: 907 corr = inex && neg; 908 break; 909 case MPFR_RNDA: 910 corr = inex && !neg; 911 break; 912 default: 913 MPFR_ASSERTN (rnd == MPFR_RNDN); 914 /* Note: for halfway cases (maxexp == MPFR_EXP_MIN) that are 915 rounded downward, rbit has been changed to 0 so that corr 916 is set correctly. */ 917 corr = rbit; 918 } 919 MPFR_ASSERTD (corr == 0 || corr == 1); 920 if (inex && 921 corr == 0) /* two's complement significand decreased */ 922 inex = -1; 923 } 924 else /* tmd */ 925 { 926 mpfr_exp_t minexp2; 927 mpfr_prec_t cancel2; 928 mpfr_exp_t err; /* exponent of the error bound */ 929 mp_size_t zz; /* nb of limbs to zero in the TMD accumulator */ 930 mp_limb_t *zp; /* pointer to the TMD accumulator */ 931 mpfr_prec_t zq; /* size of the TMD accumulator, in bits */ 932 int sst; /* sign of the secondary term */ 933 934 /* TMD case. Here we use a new variable minexp2, with the same 935 meaning as minexp, as we want to keep the minexp value for 936 the copy to the destination. */ 937 938 MPFR_ASSERTD (maxexp > MPFR_EXP_MIN); 939 MPFR_ASSERTD (tmd == 1 || tmd == 2); 940 941 /* TMD accumulator */ 942 zp = wp + ws; 943 zq = (mpfr_prec_t) zs * GMP_NUMB_BITS; 944 945 err = maxexp + logn; 946 947 MPFR_LOG_MSG (("TMD with" 948 " maxexp=%" MPFR_EXP_FSPEC "d" 949 " err=%" MPFR_EXP_FSPEC "d" 950 " zs=%Pd" 951 " zq=%Pd\n", 952 (mpfr_eexp_t) maxexp, (mpfr_eexp_t) err, 953 (mpfr_prec_t) zs, zq)); 954 955 /* The d-1 bits from u-2 to u-d (= err) are identical. */ 956 957 if (err >= minexp) 958 { 959 mpfr_prec_t tq; 960 mp_size_t wi; 961 int td; 962 963 /* Let's keep the last 2 over the d-1 identical bits and the 964 following bits, i.e. the bits from err+1 to minexp. */ 965 tq = err - minexp + 2; /* tq = number of such bits */ 966 MPFR_LOG_MSG (("[TMD] tq=%Pd\n", tq)); 967 MPFR_ASSERTD (tq >= 2); 968 969 wi = tq / GMP_NUMB_BITS; 970 td = tq % GMP_NUMB_BITS; 971 972 /* Note: The "else" (td == 0) branch below can be executed 973 only if tq >= GMP_NUMB_BITS, which is possible only when 974 logn is large enough. Indeed, if tq > logn + some constant, 975 this means that the TMD did not occur. 976 TODO: Find an upper bound on tq, and add a corresponding 977 MPFR_ASSERTD assertion / hint. On some platforms, this 978 branch might be dead code, and such information would 979 allow the compiler to remove it. 980 It seems that this branch is never tested (r12754). */ 981 982 if (td != 0) 983 { 984 wi++; /* number of words with represented bits */ 985 td = GMP_NUMB_BITS - td; 986 zz = zs - wi; 987 MPFR_ASSERTD (zz >= 0 && zz < zs); 988 mpn_lshift (zp + zz, wp, wi, td); 989 } 990 else 991 { 992 MPFR_ASSERTD (wi > 0); 993 zz = zs - wi; 994 MPFR_ASSERTD (zz >= 0 && zz < zs); 995 if (zz > 0) 996 MPN_COPY (zp + zz, wp, wi); 997 } 998 999 /* Compute minexp2 = minexp - (zs * GMP_NUMB_BITS + td) 1000 safely. */ 1001 SAFE_SUB (minexp2, minexp, zz * GMP_NUMB_BITS + td); 1002 MPFR_ASSERTD (minexp2 == err + 2 - zq); 1003 } 1004 else /* err < minexp */ 1005 { 1006 /* At least one of the identical bits is not represented, 1007 meaning that it is 0 and all these bits are 0's. Thus 1008 the accumulator will be 0. The new minexp is determined 1009 from maxexp, with cq bits reserved to avoid an overflow 1010 (as in the early steps). */ 1011 MPFR_LOG_MSG (("[TMD] err < minexp\n", 0)); 1012 zz = zs; 1013 1014 /* Compute minexp2 = maxexp - (zq - cq) safely. */ 1015 SAFE_SUB (minexp2, maxexp, zq - cq); 1016 MPFR_ASSERTD (minexp2 == err + 1 - zq); 1017 } 1018 1019 MPN_ZERO (zp, zz); 1020 1021 /* We need to determine the sign sst of the secondary term. 1022 In sum_raw, since the truncated sum corresponding to this 1023 secondary term will be in [2^(e-1),2^e] and the error 1024 strictly less than 2^err, we can stop the iterations when 1025 e - err >= 1 (this bound is the 11th argument of sum_raw). */ 1026 cancel2 = sum_raw (zp, zs, zq, x, n, minexp2, maxexp, tp, ts, 1027 logn, 1, NULL, NULL, NULL); 1028 1029 if (cancel2 != 0) 1030 sst = MPFR_LIMB_MSB (zp[zs-1]) == 0 ? 1 : -1; 1031 else if (tmd == 1) 1032 sst = 0; 1033 else 1034 { 1035 /* For halfway cases, let's virtually eliminate them 1036 by setting a sst equivalent to a non-halfway case, 1037 which depends on the last bit of the pre-rounded 1038 result. */ 1039 MPFR_ASSERTD (rnd == MPFR_RNDN && tmd == 2); 1040 sst = lbit != 0 ? 1 : -1; 1041 } 1042 1043 MPFR_LOG_MSG (("[TMD] tmd=%d rbit=%d sst=%d\n", 1044 tmd, (int) rbit, sst)); 1045 1046 /* Do not consider the corrected sst for MPFR_COV_SET */ 1047 MPFR_COV_SET (sum_tmd[(int) rnd][tmd-1][rbit] 1048 [cancel2 == 0 ? 1 : sst+1][neg][sq > MPFR_PREC_MIN]); 1049 1050 inex = 1051 MPFR_IS_LIKE_RNDD (rnd, sgn) ? (sst ? -1 : 0) : 1052 MPFR_IS_LIKE_RNDU (rnd, sgn) ? (sst ? 1 : 0) : 1053 (MPFR_ASSERTD (rnd == MPFR_RNDN), 1054 tmd == 1 ? - sst : sst); 1055 1056 if (tmd == 2 && sst == (rbit != 0 ? -1 : 1)) 1057 corr = 1 - (int) rbit; 1058 else if (MPFR_IS_LIKE_RNDD (rnd, sgn) && sst == -1) 1059 corr = (int) rbit - 1; 1060 else if (MPFR_IS_LIKE_RNDU (rnd, sgn) && sst == +1) 1061 corr = (int) rbit + 1; 1062 else 1063 corr = (int) rbit; 1064 } /* tmd */ 1065 } /* rnd != MPFR_RNDF */ 1066 1067 MPFR_LOG_MSG (("neg=%d corr=%d inex=%d\n", neg, corr, inex)); 1068 1069 /* Sign handling (-> absolute value and sign), together with 1070 rounding. The most common cases are corr = 0 and corr = 1 1071 as this is necessarily the case when the TMD did not occur. */ 1072 1073 MPFR_ASSERTD (corr >= -1 && corr <= 2); 1074 1075 MPFR_SIGN (sum) = sgn; 1076 1077 /* Let's copy/shift the bits [max(u,minexp),e) to the 1078 most significant part of the destination, and zero 1079 the least significant part (there can be one only if 1080 u < minexp). The trailing bits of the destination may 1081 contain garbage at this point. */ 1082 1083 sn = MPFR_PREC2LIMBS (sq); 1084 sd = (mpfr_prec_t) sn * GMP_NUMB_BITS - sq; 1085 sh = cancel % GMP_NUMB_BITS; 1086 1087 MPFR_ASSERTD (sd >= 0 && sd < GMP_NUMB_BITS); 1088 1089 if (MPFR_LIKELY (u > minexp)) 1090 { 1091 mp_size_t wi; 1092 1093 /* Recompute the initial value of wi. */ 1094 wi = (u - minexp) / GMP_NUMB_BITS; 1095 if (MPFR_LIKELY (sh != 0)) 1096 { 1097 mp_size_t fi; 1098 1099 fi = (e - minexp) / GMP_NUMB_BITS - (sn - 1); 1100 MPFR_ASSERTD (fi == wi || fi == wi + 1); 1101 mpn_lshift (sump, wp + fi, sn, sh); 1102 if (fi != wi) 1103 sump[0] |= wp[wi] >> (GMP_NUMB_BITS - sh); 1104 } 1105 else 1106 { 1107 MPFR_ASSERTD ((mpfr_prec_t) (ws - (wi + sn)) * GMP_NUMB_BITS 1108 == cancel); 1109 MPN_COPY (sump, wp + wi, sn); 1110 } 1111 } 1112 else /* u <= minexp */ 1113 { 1114 mp_size_t en; 1115 1116 en = (e - minexp + (GMP_NUMB_BITS - 1)) / GMP_NUMB_BITS; 1117 if (MPFR_LIKELY (sh != 0)) 1118 mpn_lshift (sump + sn - en, wp, en, sh); 1119 else if (MPFR_UNLIKELY (en > 0)) 1120 MPN_COPY (sump + sn - en, wp, en); 1121 if (sn > en) 1122 MPN_ZERO (sump, sn - en); 1123 } 1124 1125 /* Let's take the complement if the result is negative, and at 1126 the same time, do the rounding and zero the trailing bits. 1127 As this is valid only for precisions >= 2, there is special 1128 code for precision 1 first. */ 1129 1130 if (MPFR_UNLIKELY (sq == 1)) /* precision 1 */ 1131 { 1132 sump[0] = MPFR_LIMB_HIGHBIT; 1133 e += neg ? 1 - corr : corr; 1134 } 1135 else if (neg) /* negative result with sq > 1 */ 1136 { 1137 MPFR_ASSERTD (MPFR_LIMB_MSB (sump[sn-1]) == 0); 1138 1139 /* abs(x + corr) = - (x + corr) = com(x) + (1 - corr) */ 1140 1141 /* We want to avoid separate mpn_com (or mpn_neg) and mpn_add_1 1142 (or mpn_sub_1) operations, as they could yield two loops in 1143 some particular cases involving a long sequence of 0's in 1144 the low significant bits (except the least significant bit, 1145 which doesn't matter). */ 1146 1147 if (corr <= 1) 1148 { 1149 mp_limb_t corr2; 1150 1151 /* Here we can just do the correction operation on the 1152 least significant limb, then do either a mpn_com or 1153 a mpn_neg on the remaining limbs, depending on the 1154 carry (BTW, mpn_neg is just a mpn_com with an initial 1155 carry propagation: after some point, mpn_neg does a 1156 complement). */ 1157 1158 corr2 = (mp_limb_t) (1 - corr) << sd; 1159 /* Note: If corr = -1, this can overflow to corr2 = 0. 1160 This case is taken into account below. */ 1161 1162 sump[0] = (~ (sump[0] | MPFR_LIMB_MASK (sd))) + corr2; 1163 1164 if (sump[0] < corr2 || (corr2 == 0 && corr < 0)) 1165 { 1166 if (sn == 1 || ! mpn_neg (sump + 1, sump + 1, sn - 1)) 1167 { 1168 /* Note: The | is important when sump[sn-1] is not 0 1169 (this can occur with sn = 1 and corr = -1). TODO: 1170 Add something to make sure that this is tested. */ 1171 sump[sn-1] |= MPFR_LIMB_HIGHBIT; 1172 e++; 1173 } 1174 } 1175 else if (sn > 1) 1176 mpn_com (sump + 1, sump + 1, sn - 1); 1177 } 1178 else /* corr == 2 */ 1179 { 1180 mp_limb_t corr2, c; 1181 mp_size_t i = 1; 1182 1183 /* We want to compute com(x) - 1, but GMP doesn't have an 1184 operation for that. The fact is that a sequence of low 1185 significant bits 1 is invariant. Starting at the first 1186 low significant bit 0, we can do the complement with 1187 mpn_com. */ 1188 1189 corr2 = MPFR_LIMB_ONE << sd; 1190 c = ~ (sump[0] | MPFR_LIMB_MASK (sd)); 1191 sump[0] = c - corr2; 1192 1193 if (c == 0) 1194 { 1195 while (MPFR_ASSERTD (i < sn), sump[i] == MPFR_LIMB_MAX) 1196 i++; 1197 sump[i] = (~ sump[i]) - 1; 1198 i++; 1199 } 1200 1201 if (i < sn) 1202 mpn_com (sump + i, sump + i, sn - i); 1203 else if (MPFR_UNLIKELY (MPFR_LIMB_MSB (sump[sn-1]) == 0)) 1204 { 1205 /* Happens on 01111...111, whose complement is 1206 10000...000, and com(x) - 1 is 01111...111. */ 1207 sump[sn-1] |= MPFR_LIMB_HIGHBIT; 1208 e--; 1209 } 1210 } 1211 } 1212 else /* positive result with sq > 1 */ 1213 { 1214 MPFR_ASSERTD (MPFR_LIMB_MSB (sump[sn-1]) != 0); 1215 sump[0] &= ~ MPFR_LIMB_MASK (sd); 1216 1217 if (corr > 0) 1218 { 1219 mp_limb_t corr2, carry_out; 1220 1221 corr2 = (mp_limb_t) corr << sd; 1222 /* If corr == 2 && sd == GMP_NUMB_BITS - 1, this overflows 1223 to corr2 = 0. This case is taken into account below. */ 1224 1225 carry_out = corr2 != 0 ? mpn_add_1 (sump, sump, sn, corr2) : 1226 (MPFR_ASSERTD (sn > 1), 1227 mpn_add_1 (sump + 1, sump + 1, sn - 1, MPFR_LIMB_ONE)); 1228 1229 MPFR_ASSERTD (sump[sn-1] >> (GMP_NUMB_BITS - 1) == !carry_out); 1230 1231 if (MPFR_UNLIKELY (carry_out)) 1232 { 1233 /* Note: The | is important when sump[sn-1] is not 0 1234 (this can occur with sn = 1 and corr = 2). TODO: 1235 Add something to make sure that this is tested. */ 1236 sump[sn-1] |= MPFR_LIMB_HIGHBIT; 1237 e++; 1238 } 1239 } 1240 1241 if (corr < 0) 1242 { 1243 mpn_sub_1 (sump, sump, sn, MPFR_LIMB_ONE << sd); 1244 1245 if (MPFR_UNLIKELY (MPFR_LIMB_MSB (sump[sn-1]) == 0)) 1246 { 1247 sump[sn-1] |= MPFR_LIMB_HIGHBIT; 1248 e--; 1249 } 1250 } 1251 } 1252 1253 MPFR_ASSERTD (MPFR_LIMB_MSB (sump[sn-1]) != 0); 1254 MPFR_LOG_MSG (("Set exponent e=%" MPFR_EXP_FSPEC "d\n", (mpfr_eexp_t) e)); 1255 /* e may be outside the current exponent range, but this will be checked 1256 with mpfr_check_range below. */ 1257 MPFR_EXP (sum) = e; 1258 } /* main block */ 1259 1260 MPFR_TMP_FREE (marker); 1261 return mpfr_check_range (sum, inex, rnd); 1262 } 1263 1264 /**********************************************************************/ 1265 1266 int 1267 mpfr_sum (mpfr_ptr sum, const mpfr_ptr *x, unsigned long n, mpfr_rnd_t rnd) 1268 { 1269 MPFR_LOG_FUNC 1270 (("n=%lu rnd=%d", n, rnd), 1271 ("sum[%Pd]=%.*Rg", mpfr_get_prec (sum), mpfr_log_prec, sum)); 1272 1273 if (MPFR_UNLIKELY (n <= 2)) 1274 { 1275 if (n == 0) 1276 { 1277 MPFR_SET_ZERO (sum); 1278 MPFR_SET_POS (sum); 1279 MPFR_RET (0); 1280 } 1281 else if (n == 1) 1282 return mpfr_set (sum, x[0], rnd); 1283 else 1284 return mpfr_add (sum, x[0], x[1], rnd); 1285 } 1286 else 1287 { 1288 mpfr_exp_t maxexp = MPFR_EXP_MIN; /* max(Empty) */ 1289 unsigned long i; 1290 unsigned long rn = 0; /* will be the number of regular inputs */ 1291 /* sign of infinities and zeros (0: currently unknown) */ 1292 int sign_inf = 0, sign_zero = 0; 1293 1294 MPFR_LOG_MSG (("Check for special inputs (n = %lu >= 3)\n", n)); 1295 1296 for (i = 0; i < n; i++) 1297 { 1298 if (MPFR_UNLIKELY (MPFR_IS_SINGULAR (x[i]))) 1299 { 1300 if (MPFR_IS_NAN (x[i])) 1301 { 1302 /* The current value x[i] is NaN. Then the sum is NaN. */ 1303 nan: 1304 MPFR_SET_NAN (sum); 1305 MPFR_RET_NAN; 1306 } 1307 else if (MPFR_IS_INF (x[i])) 1308 { 1309 /* The current value x[i] is an infinity. 1310 There are two cases: 1311 1. This is the first infinity value (sign_inf == 0). 1312 Then set sign_inf to its sign, and go on. 1313 2. All the infinities found until now have the same 1314 sign sign_inf. If this new infinity has a different 1315 sign, then return NaN immediately, else go on. */ 1316 if (sign_inf == 0) 1317 sign_inf = MPFR_SIGN (x[i]); 1318 else if (MPFR_SIGN (x[i]) != sign_inf) 1319 goto nan; 1320 } 1321 else if (MPFR_UNLIKELY (rn == 0)) 1322 { 1323 /* The current value x[i] is a zero. The code below matters 1324 only when all values found until now are zeros, otherwise 1325 it is harmless (the test rn == 0 above is just a minor 1326 optimization). 1327 Here we track the sign of the zero result when all inputs 1328 are zeros: if all zeros have the same sign, the result 1329 will have this sign, otherwise (i.e. if there is at least 1330 a zero of each sign), the sign of the zero result depends 1331 only on the rounding mode (note that this choice is 1332 sticky when new zeros are considered). */ 1333 MPFR_ASSERTD (MPFR_IS_ZERO (x[i])); 1334 if (sign_zero == 0) 1335 sign_zero = MPFR_SIGN (x[i]); 1336 else if (MPFR_SIGN (x[i]) != sign_zero) 1337 sign_zero = rnd == MPFR_RNDD ? -1 : 1; 1338 } 1339 } 1340 else 1341 { 1342 /* The current value x[i] is a regular number. */ 1343 mpfr_exp_t e = MPFR_GET_EXP (x[i]); 1344 if (e > maxexp) 1345 maxexp = e; /* maximum exponent found until now */ 1346 rn++; /* current number of regular inputs */ 1347 } 1348 } 1349 1350 MPFR_LOG_MSG (("rn=%lu sign_inf=%d sign_zero=%d\n", 1351 rn, sign_inf, sign_zero)); 1352 1353 /* At this point the result cannot be NaN (this case has already 1354 been filtered out). */ 1355 1356 if (MPFR_UNLIKELY (sign_inf != 0)) 1357 { 1358 /* At least one infinity, and all of them have the same sign 1359 sign_inf. The sum is the infinity of this sign. */ 1360 MPFR_SET_INF (sum); 1361 MPFR_SET_SIGN (sum, sign_inf); 1362 MPFR_RET (0); 1363 } 1364 1365 /* At this point, all the inputs are finite numbers. */ 1366 1367 if (MPFR_UNLIKELY (rn == 0)) 1368 { 1369 /* All the numbers were zeros (and there is at least one). 1370 The sum is zero with sign sign_zero. */ 1371 MPFR_ASSERTD (sign_zero != 0); 1372 MPFR_SET_ZERO (sum); 1373 MPFR_SET_SIGN (sum, sign_zero); 1374 MPFR_RET (0); 1375 } 1376 1377 /* Optimize the case where there are only two regular numbers. */ 1378 if (MPFR_UNLIKELY (rn <= 2)) 1379 { 1380 unsigned long h = ULONG_MAX; 1381 1382 for (i = 0; i < n; i++) 1383 if (! MPFR_IS_SINGULAR (x[i])) 1384 { 1385 if (rn == 1) 1386 return mpfr_set (sum, x[i], rnd); 1387 if (h != ULONG_MAX) 1388 return mpfr_add (sum, x[h], x[i], rnd); 1389 h = i; 1390 } 1391 MPFR_RET_NEVER_GO_HERE(); 1392 } 1393 1394 return sum_aux (sum, x, n, rnd, maxexp, rn); 1395 } 1396 } 1397
Indexes created Fri Jun 05 00:26:10 UTC 2026