| 1 | /* |
| 2 | * This Source Code Form is subject to the terms of the Mozilla Public |
| 3 | * License, v. 2.0. If a copy of the MPL was not distributed with this |
| 4 | * file, You can obtain one at http://mozilla.org/MPL/2.0/. |
| 5 | * |
| 6 | * Copyright 1997 - July 2008 CWI, August 2008 - 2019 MonetDB B.V. |
| 7 | */ |
| 8 | |
| 9 | /* |
| 10 | * This file implements a stable sort algorithm known as "timsort". |
| 11 | * The algorithm is a straight copy of the listsort function in the |
| 12 | * Python 2.5 source code, heavily modified to fit into the MonetDB |
| 13 | * environment. |
| 14 | * The original author of the sort algorithm was Tim Peters, the |
| 15 | * adaptation was done by Sjoerd Mullender. |
| 16 | * |
| 17 | * |
| 18 | * This file is included multiple times. We expect a bunch of tokens |
| 19 | * to be redefined differently each time (see gdk_ssort.c). If the |
| 20 | * token GDKssortimpl is defined, the main interface is defined. |
| 21 | */ |
| 22 | |
| 23 | /* binarysort is the best method for sorting small arrays: it does few |
| 24 | * compares, but can do data movement quadratic in the number of |
| 25 | * elements. |
| 26 | * [lo, hi) is a contiguous slice of a list, and is sorted via binary |
| 27 | * insertion. This sort is stable. On entry, must have lo <= start <= |
| 28 | * hi, and that [lo, start) is already sorted (pass start == lo if you |
| 29 | * don't know!). */ |
| 30 | static void |
| 31 | binarysort(size_t lo, size_t hi, size_t start, MergeState *ms) |
| 32 | { |
| 33 | register size_t l, p, r; |
| 34 | |
| 35 | assert(lo <= start && start <= hi); |
| 36 | /* assert [lo, start) is sorted */ |
| 37 | if (lo == start) |
| 38 | start++; |
| 39 | /* [lo,start) is sorted, insert start (the pivot) into this |
| 40 | * area, finding its position using binary search. */ |
| 41 | for (; start < hi; start++) { |
| 42 | /* set l to where start belongs */ |
| 43 | l = lo; |
| 44 | r = start; |
| 45 | /* ms->t[ht] is the pivot */ |
| 46 | COPY(ms->th, PTRADD(ms->bh, r, ms->hs), ms->hs); |
| 47 | COPY_any(ms->tt, PTRADD(ms->bt, r, ms->ts), ms->ts); |
| 48 | /* Invariants: |
| 49 | * pivot >= all in [lo, l). |
| 50 | * pivot < all in [r, start). |
| 51 | * The second is vacuously true at the start. */ |
| 52 | assert(l < r); |
| 53 | do { |
| 54 | p = l + ((r - l) >> 1); |
| 55 | if (ISLT(ms->th, PTRADD(ms->bh, p, ms->hs), ms)) |
| 56 | r = p; |
| 57 | else |
| 58 | l = p + 1; |
| 59 | } while (l < r); |
| 60 | assert(l == r); |
| 61 | /* The invariants still hold, so pivot >= all in [lo, |
| 62 | * l) and pivot < all in [l, start), so pivot belongs |
| 63 | * at l. Note that if there are elements equal to |
| 64 | * pivot, l points to the first slot after them -- |
| 65 | * that's why this sort is stable. Slide over to make |
| 66 | * room. |
| 67 | * Caution: using memmove is much slower under MSVC 5; |
| 68 | * we're not usually moving many slots. */ |
| 69 | for (p = start, r = p - 1; p > l; p = r, r = p - 1) { |
| 70 | COPY(PTRADD(ms->bh, p, ms->hs), |
| 71 | PTRADD(ms->bh, r, ms->hs), ms->hs); |
| 72 | COPY_any(PTRADD(ms->bt, p, ms->ts), |
| 73 | PTRADD(ms->bt, r, ms->ts), ms->ts); |
| 74 | } |
| 75 | COPY(PTRADD(ms->bh, l, ms->hs), ms->th, ms->hs); |
| 76 | COPY_any(PTRADD(ms->bt, l, ms->ts), ms->tt, ms->ts); |
| 77 | } |
| 78 | } |
| 79 | |
| 80 | /* Locate the proper position of key in a sorted vector; if the vector |
| 81 | * contains an element equal to key, return the position immediately |
| 82 | * to the left of the leftmost equal element. [gallop_right() does |
| 83 | * the same except returns the position to the right of the rightmost |
| 84 | * equal element (if any).] |
| 85 | * |
| 86 | * "a" is a sorted vector with n elements, starting at a[0]. n must |
| 87 | * be > 0. |
| 88 | * |
| 89 | * "hint" is an index at which to begin the search, 0 <= hint < n. |
| 90 | * The closer hint is to the final result, the faster this runs. |
| 91 | * |
| 92 | * The return value is the int k in 0..n such that |
| 93 | * |
| 94 | * a[k-1] < key <= a[k] |
| 95 | * |
| 96 | * pretending that *(a-1) is minus infinity and a[n] is plus infinity. |
| 97 | * IOW, key belongs at index k; or, IOW, the first k elements of a |
| 98 | * should precede key, and the last n-k should follow key. |
| 99 | * |
| 100 | * Returns -1 on error. See listsort.txt for info on the method. */ |
| 101 | static ssize_t |
| 102 | gallop_left(void *key, void *a, ssize_t n, ssize_t hint, MergeState *ms) |
| 103 | { |
| 104 | ssize_t ofs; |
| 105 | ssize_t lastofs; |
| 106 | ssize_t k; |
| 107 | |
| 108 | assert(key && a && n > 0 && hint >= 0 && hint < n); |
| 109 | |
| 110 | a = PTRADD(a, hint, ms->hs); |
| 111 | lastofs = 0; |
| 112 | ofs = 1; |
| 113 | if (ISLT(a, key, ms)) { |
| 114 | /* a[hint] < key -- gallop right, until |
| 115 | * a[hint + lastofs] < key <= a[hint + ofs] */ |
| 116 | const ssize_t maxofs = n - hint; /* &a[n-1] is highest */ |
| 117 | while (ofs < maxofs) { |
| 118 | if (ISLT(PTRADD(a, ofs, ms->hs), key, ms)) { |
| 119 | lastofs = ofs; |
| 120 | ofs = (ofs << 1) + 1; |
| 121 | if (ofs <= 0) /* int overflow */ |
| 122 | ofs = maxofs; |
| 123 | } else /* key <= a[hint + ofs] */ |
| 124 | break; |
| 125 | } |
| 126 | if (ofs > maxofs) |
| 127 | ofs = maxofs; |
| 128 | /* Translate back to offsets relative to &a[0]. */ |
| 129 | lastofs += hint; |
| 130 | ofs += hint; |
| 131 | } else { |
| 132 | /* key <= a[hint] -- gallop left, until |
| 133 | * a[hint - ofs] < key <= a[hint - lastofs] */ |
| 134 | const ssize_t maxofs = hint + 1; /* &a[0] is lowest */ |
| 135 | while (ofs < maxofs) { |
| 136 | if (ISLT(PTRADD(a, -ofs, ms->hs), key, ms)) |
| 137 | break; |
| 138 | /* key <= a[hint - ofs] */ |
| 139 | lastofs = ofs; |
| 140 | ofs = (ofs << 1) + 1; |
| 141 | if (ofs <= 0) /* int overflow */ |
| 142 | ofs = maxofs; |
| 143 | } |
| 144 | if (ofs > maxofs) |
| 145 | ofs = maxofs; |
| 146 | /* Translate back to positive offsets relative to &a[0]. */ |
| 147 | k = lastofs; |
| 148 | lastofs = hint - ofs; |
| 149 | ofs = hint - k; |
| 150 | } |
| 151 | a = PTRADD(a, -hint, ms->hs); |
| 152 | |
| 153 | assert(-1 <= lastofs && lastofs < ofs && ofs <= n); |
| 154 | /* Now a[lastofs] < key <= a[ofs], so key belongs somewhere to |
| 155 | * the right of lastofs but no farther right than ofs. Do a |
| 156 | * binary search, with invariant a[lastofs-1] < key <= |
| 157 | * a[ofs]. */ |
| 158 | ++lastofs; |
| 159 | while (lastofs < ofs) { |
| 160 | ssize_t m = lastofs + ((ofs - lastofs) >> 1); |
| 161 | |
| 162 | if (ISLT(PTRADD(a, m, ms->hs), key, ms)) |
| 163 | lastofs = m + 1; /* a[m] < key */ |
| 164 | else |
| 165 | ofs = m; /* key <= a[m] */ |
| 166 | } |
| 167 | assert(lastofs == ofs); /* so a[ofs-1] < key <= a[ofs] */ |
| 168 | return ofs; |
| 169 | } |
| 170 | |
| 171 | /* Exactly like gallop_left(), except that if key already exists in |
| 172 | * a[0:n], finds the position immediately to the right of the |
| 173 | * rightmost equal value. |
| 174 | * |
| 175 | * The return value is the int k in 0..n such that |
| 176 | * |
| 177 | * a[k-1] <= key < a[k] |
| 178 | * |
| 179 | * or -1 if error. |
| 180 | * |
| 181 | * The code duplication is massive, but this is enough different given |
| 182 | * that we're sticking to "<" comparisons that it's much harder to |
| 183 | * follow if written as one routine with yet another "left or right?" |
| 184 | * flag. */ |
| 185 | static ssize_t |
| 186 | gallop_right(void *key, void *a, ssize_t n, ssize_t hint, MergeState *ms) |
| 187 | { |
| 188 | ssize_t ofs; |
| 189 | ssize_t lastofs; |
| 190 | ssize_t k; |
| 191 | |
| 192 | assert(key && a && n > 0 && hint >= 0 && hint < n); |
| 193 | |
| 194 | a = PTRADD(a, hint, ms->hs); |
| 195 | lastofs = 0; |
| 196 | ofs = 1; |
| 197 | if (ISLT(key, a, ms)) { |
| 198 | /* key < a[hint] -- gallop left, until |
| 199 | * a[hint - ofs] <= key < a[hint - lastofs] */ |
| 200 | const ssize_t maxofs = hint + 1; /* &a[0] is lowest */ |
| 201 | while (ofs < maxofs) { |
| 202 | if (ISLT(key, PTRADD(a, -ofs, ms->hs), ms)) { |
| 203 | lastofs = ofs; |
| 204 | ofs = (ofs << 1) + 1; |
| 205 | if (ofs <= 0) /* int overflow */ |
| 206 | ofs = maxofs; |
| 207 | } else /* a[hint - ofs] <= key */ |
| 208 | break; |
| 209 | } |
| 210 | if (ofs > maxofs) |
| 211 | ofs = maxofs; |
| 212 | /* Translate back to positive offsets relative to &a[0]. */ |
| 213 | k = lastofs; |
| 214 | lastofs = hint - ofs; |
| 215 | ofs = hint - k; |
| 216 | } else { |
| 217 | /* a[hint] <= key -- gallop right, until |
| 218 | * a[hint + lastofs] <= key < a[hint + ofs] */ |
| 219 | const ssize_t maxofs = n - hint; /* &a[n-1] is highest */ |
| 220 | while (ofs < maxofs) { |
| 221 | if (ISLT(key, PTRADD(a, ofs, ms->hs), ms)) |
| 222 | break; |
| 223 | /* a[hint + ofs] <= key */ |
| 224 | lastofs = ofs; |
| 225 | ofs = (ofs << 1) + 1; |
| 226 | if (ofs <= 0) /* int overflow */ |
| 227 | ofs = maxofs; |
| 228 | } |
| 229 | if (ofs > maxofs) |
| 230 | ofs = maxofs; |
| 231 | /* Translate back to offsets relative to &a[0]. */ |
| 232 | lastofs += hint; |
| 233 | ofs += hint; |
| 234 | } |
| 235 | a = PTRADD(a, -hint, ms->hs); |
| 236 | |
| 237 | assert(-1 <= lastofs && lastofs < ofs && ofs <= n); |
| 238 | /* Now a[lastofs] <= key < a[ofs], so key belongs somewhere to |
| 239 | * the right of lastofs but no farther right than ofs. Do a |
| 240 | * binary search, with invariant a[lastofs-1] <= key < |
| 241 | * a[ofs]. */ |
| 242 | ++lastofs; |
| 243 | while (lastofs < ofs) { |
| 244 | ssize_t m = lastofs + ((ofs - lastofs) >> 1); |
| 245 | |
| 246 | if (ISLT(key, PTRADD(a, m, ms->hs), ms)) |
| 247 | ofs = m; /* key < a[m] */ |
| 248 | else |
| 249 | lastofs = m+1; /* a[m] <= key */ |
| 250 | } |
| 251 | assert(lastofs == ofs); /* so a[ofs-1] <= key < a[ofs] */ |
| 252 | return ofs; |
| 253 | } |
| 254 | |
| 255 | /* Merge the two runs at stack indices i and i+1. |
| 256 | * Returns 0 on success, -1 on error. */ |
| 257 | static ssize_t |
| 258 | merge_at(MergeState *ms, ssize_t i) |
| 259 | { |
| 260 | size_t pa, pb; |
| 261 | ssize_t na, nb; |
| 262 | ssize_t k; |
| 263 | |
| 264 | assert(ms != NULL); |
| 265 | assert(ms->n >= 2); |
| 266 | assert(i >= 0); |
| 267 | assert(i == ms->n - 2 || i == ms->n - 3); |
| 268 | |
| 269 | pa = ms->pending[i].base; |
| 270 | na = ms->pending[i].len; |
| 271 | pb = ms->pending[i + 1].base; |
| 272 | nb = ms->pending[i + 1].len; |
| 273 | assert(na > 0 && nb > 0); |
| 274 | assert(pa + na == pb); |
| 275 | |
| 276 | /* Record the length of the combined runs; if i is the |
| 277 | * 3rd-last run now, also slide over the last run (which isn't |
| 278 | * involved in this merge). The current run i+1 goes away in |
| 279 | * any case. */ |
| 280 | ms->pending[i].len = na + nb; |
| 281 | if (i == ms->n - 3) |
| 282 | ms->pending[i + 1] = ms->pending[i + 2]; |
| 283 | --ms->n; |
| 284 | |
| 285 | /* Where does b start in a? Elements in a before that can be |
| 286 | * ignored (already in place). */ |
| 287 | k = gallop_right(PTRADD(ms->bh, pb, ms->hs), |
| 288 | PTRADD(ms->bh, pa, ms->hs), na, 0, ms); |
| 289 | pa += k; |
| 290 | na -= k; |
| 291 | if (na == 0) |
| 292 | return 0; |
| 293 | |
| 294 | /* Where does a end in b? Elements in b after that can be |
| 295 | * ignored (already in place). */ |
| 296 | nb = gallop_left(PTRADD(ms->bh, pa + na - 1, ms->hs), |
| 297 | PTRADD(ms->bh, pb, ms->hs), nb, nb-1, ms); |
| 298 | if (nb <= 0) |
| 299 | return nb; |
| 300 | |
| 301 | /* Merge what remains of the runs, using a temp array with |
| 302 | * min(na, nb) elements. */ |
| 303 | if (na <= nb) { |
| 304 | /* Merge the na elements starting at pa with the nb elements starting |
| 305 | * at pb in a stable way, in-place. na and nb must be > 0, and pa + |
| 306 | * na == pb. Must also have that *pb < *pa, that pa[na-1] belongs at |
| 307 | * the end of the merge, and should have na <= nb. See listsort.txt |
| 308 | * for more info. Return 0 if successful, -1 if error. */ |
| 309 | size_t dest; |
| 310 | ssize_t min_gallop = ms->min_gallop; |
| 311 | |
| 312 | assert(ms && na > 0 && nb > 0 && pa + na == pb); |
| 313 | if (MERGE_GETMEMH(ms, na) < 0) |
| 314 | return -1; |
| 315 | if (MERGE_GETMEMT(ms, na) < 0) |
| 316 | return -1; |
| 317 | COPY_anyN(ms->ah, PTRADD(ms->bh, pa, ms->hs), ms->hs, na); |
| 318 | COPY_anyN(ms->at, PTRADD(ms->bt, pa, ms->ts), ms->ts, na); |
| 319 | dest = pa; |
| 320 | pa = 0; |
| 321 | |
| 322 | COPY(PTRADD(ms->bh, dest, ms->hs), |
| 323 | PTRADD(ms->bh, pb, ms->hs), ms->hs); |
| 324 | COPY_any(PTRADD(ms->bt, dest, ms->ts), |
| 325 | PTRADD(ms->bt, pb, ms->ts), ms->ts); |
| 326 | dest++; |
| 327 | pb++; |
| 328 | --nb; |
| 329 | if (nb == 0) |
| 330 | goto SucceedA; |
| 331 | if (na == 1) |
| 332 | goto CopyB; |
| 333 | |
| 334 | for (;;) { |
| 335 | ssize_t acount = 0; /* # of times A won in a row */ |
| 336 | ssize_t bcount = 0; /* # of times B won in a row */ |
| 337 | |
| 338 | /* Do the straightforward thing until (if |
| 339 | * ever) one run appears to win |
| 340 | * consistently. */ |
| 341 | for (;;) { |
| 342 | assert(na > 1 && nb > 0); |
| 343 | k = ISLT(PTRADD(ms->bh, pb, ms->hs), |
| 344 | PTRADD(ms->ah, pa, ms->hs), ms); |
| 345 | if (k) { |
| 346 | COPY(PTRADD(ms->bh, dest, ms->hs), |
| 347 | PTRADD(ms->bh, pb, ms->hs), |
| 348 | ms->hs); |
| 349 | COPY_any(PTRADD(ms->bt, dest, ms->ts), |
| 350 | PTRADD(ms->bt, pb, ms->ts), |
| 351 | ms->ts); |
| 352 | dest++; |
| 353 | pb++; |
| 354 | ++bcount; |
| 355 | acount = 0; |
| 356 | --nb; |
| 357 | if (nb == 0) |
| 358 | goto SucceedA; |
| 359 | if (bcount >= min_gallop) |
| 360 | break; |
| 361 | } else { |
| 362 | COPY(PTRADD(ms->bh, dest, ms->hs), |
| 363 | PTRADD(ms->ah, pa, ms->hs), |
| 364 | ms->hs); |
| 365 | COPY_any(PTRADD(ms->bt, dest, ms->ts), |
| 366 | PTRADD(ms->at, pa, ms->ts), |
| 367 | ms->ts); |
| 368 | dest++; |
| 369 | pa++; |
| 370 | ++acount; |
| 371 | bcount = 0; |
| 372 | --na; |
| 373 | if (na == 1) |
| 374 | goto CopyB; |
| 375 | if (acount >= min_gallop) |
| 376 | break; |
| 377 | } |
| 378 | } |
| 379 | |
| 380 | /* One run is winning so consistently that |
| 381 | * galloping may be a huge win. So try that, |
| 382 | * and continue galloping until (if ever) |
| 383 | * neither run appears to be winning |
| 384 | * consistently anymore. */ |
| 385 | ++min_gallop; |
| 386 | do { |
| 387 | assert(na > 1 && nb > 0); |
| 388 | min_gallop -= min_gallop > 1; |
| 389 | ms->min_gallop = min_gallop; |
| 390 | k = gallop_right(PTRADD(ms->bh, pb, ms->hs), |
| 391 | PTRADD(ms->ah, pa, ms->hs), |
| 392 | na, 0, ms); |
| 393 | acount = k; |
| 394 | if (k) { |
| 395 | COPY_anyN(PTRADD(ms->bh, dest, ms->hs), |
| 396 | PTRADD(ms->ah, pa, ms->hs), |
| 397 | ms->hs, k); |
| 398 | COPY_anyN(PTRADD(ms->bt, dest, ms->ts), |
| 399 | PTRADD(ms->at, pa, ms->ts), |
| 400 | ms->ts, k); |
| 401 | dest += k; |
| 402 | pa += k; |
| 403 | na -= k; |
| 404 | if (na == 1) |
| 405 | goto CopyB; |
| 406 | /* na==0 is impossible now if |
| 407 | * the comparison function is |
| 408 | * consistent, but we can't |
| 409 | * assume that it is. */ |
| 410 | if (na == 0) |
| 411 | goto SucceedA; |
| 412 | } |
| 413 | COPY(PTRADD(ms->bh, dest, ms->hs), |
| 414 | PTRADD(ms->bh, pb, ms->hs), ms->hs); |
| 415 | COPY_any(PTRADD(ms->bt, dest, ms->ts), |
| 416 | PTRADD(ms->bt, pb, ms->ts), ms->ts); |
| 417 | dest++; |
| 418 | pb++; |
| 419 | --nb; |
| 420 | if (nb == 0) |
| 421 | goto SucceedA; |
| 422 | |
| 423 | k = gallop_left(PTRADD(ms->ah, pa, ms->hs), |
| 424 | PTRADD(ms->bh, pb, ms->hs), |
| 425 | nb, 0, ms); |
| 426 | bcount = k; |
| 427 | if (k) { |
| 428 | memmove(PTRADD(ms->bh, dest, ms->hs), |
| 429 | PTRADD(ms->bh, pb, ms->hs), |
| 430 | k * ms->hs); |
| 431 | memmove(PTRADD(ms->bt, dest, ms->ts), |
| 432 | PTRADD(ms->bt, pb, ms->ts), |
| 433 | k * ms->ts); |
| 434 | dest += k; |
| 435 | pb += k; |
| 436 | nb -= k; |
| 437 | if (nb == 0) |
| 438 | goto SucceedA; |
| 439 | } |
| 440 | COPY(PTRADD(ms->bh, dest, ms->hs), |
| 441 | PTRADD(ms->ah, pa, ms->hs), ms->hs); |
| 442 | COPY_any(PTRADD(ms->bt, dest, ms->ts), |
| 443 | PTRADD(ms->at, pa, ms->ts), ms->ts); |
| 444 | dest++; |
| 445 | pa++; |
| 446 | --na; |
| 447 | if (na == 1) |
| 448 | goto CopyB; |
| 449 | } while (acount >= MIN_GALLOP || bcount >= MIN_GALLOP); |
| 450 | ++min_gallop; /* penalize it for leaving galloping mode */ |
| 451 | ms->min_gallop = min_gallop; |
| 452 | } |
| 453 | SucceedA: |
| 454 | if (na) { |
| 455 | COPY_anyN(PTRADD(ms->bh, dest, ms->hs), |
| 456 | PTRADD(ms->ah, pa, ms->hs), ms->hs, na); |
| 457 | COPY_anyN(PTRADD(ms->bt, dest, ms->ts), |
| 458 | PTRADD(ms->at, pa, ms->ts), ms->ts, na); |
| 459 | } |
| 460 | return 0; |
| 461 | CopyB: |
| 462 | assert(na == 1 && nb > 0); |
| 463 | /* The last element of pa belongs at the end of the merge. */ |
| 464 | memmove(PTRADD(ms->bh, dest, ms->hs), |
| 465 | PTRADD(ms->bh, pb, ms->hs), nb * ms->hs); |
| 466 | memmove(PTRADD(ms->bt, dest, ms->ts), |
| 467 | PTRADD(ms->bt, pb, ms->ts), nb * ms->ts); |
| 468 | COPY(PTRADD(ms->bh, dest + nb, ms->hs), |
| 469 | PTRADD(ms->ah, pa, ms->hs), ms->hs); |
| 470 | COPY_any(PTRADD(ms->bt, dest + nb, ms->ts), |
| 471 | PTRADD(ms->at, pa, ms->ts), ms->ts); |
| 472 | return 0; |
| 473 | } else { |
| 474 | /* Merge the na elements starting at pa with the nb elements starting |
| 475 | * at pb in a stable way, in-place. na and nb must be > 0, and pa + |
| 476 | * na == pb. Must also have that *pb < *pa, that pa[na-1] belongs at |
| 477 | * the end of the merge, and should have na >= nb. See listsort.txt |
| 478 | * for more info. Return 0 if successful, -1 if error. */ |
| 479 | size_t dest; |
| 480 | size_t basea; |
| 481 | size_t baseb; |
| 482 | ssize_t min_gallop = ms->min_gallop; |
| 483 | |
| 484 | assert(ms && na > 0 && nb > 0 && pa + na == pb); |
| 485 | if (MERGE_GETMEMH(ms, nb) < 0) |
| 486 | return -1; |
| 487 | if (MERGE_GETMEMT(ms, nb) < 0) |
| 488 | return -1; |
| 489 | dest = pb + nb - 1; |
| 490 | COPY_anyN(ms->ah, PTRADD(ms->bh, pb, ms->hs), ms->hs, nb); |
| 491 | COPY_anyN(ms->at, PTRADD(ms->bt, pb, ms->ts), ms->ts, nb); |
| 492 | basea = pa; |
| 493 | baseb = 0; |
| 494 | pb = nb - 1; |
| 495 | pa += na - 1; |
| 496 | |
| 497 | COPY(PTRADD(ms->bh, dest, ms->hs), |
| 498 | PTRADD(ms->bh, pa, ms->hs), ms->hs); |
| 499 | COPY_any(PTRADD(ms->bt, dest, ms->ts), |
| 500 | PTRADD(ms->bt, pa, ms->ts), ms->ts); |
| 501 | dest--; |
| 502 | pa--; |
| 503 | --na; |
| 504 | if (na == 0) |
| 505 | goto SucceedB; |
| 506 | if (nb == 1) |
| 507 | goto CopyA; |
| 508 | |
| 509 | for (;;) { |
| 510 | ssize_t acount = 0; /* # of times A won in a row */ |
| 511 | ssize_t bcount = 0; /* # of times B won in a row */ |
| 512 | |
| 513 | /* Do the straightforward thing until (if |
| 514 | * ever) one run appears to win |
| 515 | * consistently. */ |
| 516 | for (;;) { |
| 517 | assert(na > 0 && nb > 1); |
| 518 | k = ISLT(PTRADD(ms->ah, pb, ms->hs), |
| 519 | PTRADD(ms->bh, pa, ms->hs), ms); |
| 520 | if (k) { |
| 521 | COPY(PTRADD(ms->bh, dest, ms->hs), |
| 522 | PTRADD(ms->bh, pa, ms->hs), |
| 523 | ms->hs); |
| 524 | COPY_any(PTRADD(ms->bt, dest, ms->ts), |
| 525 | PTRADD(ms->bt, pa, ms->ts), |
| 526 | ms->ts); |
| 527 | dest--; |
| 528 | pa--; |
| 529 | ++acount; |
| 530 | bcount = 0; |
| 531 | --na; |
| 532 | if (na == 0) |
| 533 | goto SucceedB; |
| 534 | if (acount >= min_gallop) |
| 535 | break; |
| 536 | } else { |
| 537 | COPY(PTRADD(ms->bh, dest, ms->hs), |
| 538 | PTRADD(ms->ah, pb, ms->hs), |
| 539 | ms->hs); |
| 540 | COPY_any(PTRADD(ms->bt, dest, ms->ts), |
| 541 | PTRADD(ms->at, pb, ms->ts), |
| 542 | ms->ts); |
| 543 | dest--; |
| 544 | pb--; |
| 545 | ++bcount; |
| 546 | acount = 0; |
| 547 | --nb; |
| 548 | if (nb == 1) |
| 549 | goto CopyA; |
| 550 | if (bcount >= min_gallop) |
| 551 | break; |
| 552 | } |
| 553 | } |
| 554 | |
| 555 | /* One run is winning so consistently that |
| 556 | * galloping may be a huge win. So try that, |
| 557 | * and continue galloping until (if ever) |
| 558 | * neither run appears to be winning |
| 559 | * consistently anymore. */ |
| 560 | ++min_gallop; |
| 561 | do { |
| 562 | assert(na > 0 && nb > 1); |
| 563 | min_gallop -= min_gallop > 1; |
| 564 | ms->min_gallop = min_gallop; |
| 565 | k = gallop_right(PTRADD(ms->ah, pb, ms->hs), |
| 566 | PTRADD(ms->bh, basea, ms->hs), |
| 567 | na, na - 1, ms); |
| 568 | k = na - k; |
| 569 | acount = k; |
| 570 | if (k) { |
| 571 | dest -= k; |
| 572 | pa -= k; |
| 573 | memmove(PTRADD(ms->bh, dest + 1, |
| 574 | ms->hs), |
| 575 | PTRADD(ms->bh, pa + 1, ms->hs), |
| 576 | k * ms->hs); |
| 577 | memmove(PTRADD(ms->bt, dest + 1, |
| 578 | ms->ts), |
| 579 | PTRADD(ms->bt, pa + 1, ms->ts), |
| 580 | k * ms->ts); |
| 581 | na -= k; |
| 582 | if (na == 0) |
| 583 | goto SucceedB; |
| 584 | } |
| 585 | COPY(PTRADD(ms->bh, dest, ms->hs), |
| 586 | PTRADD(ms->ah, pb, ms->hs), ms->hs); |
| 587 | COPY_any(PTRADD(ms->bt, dest, ms->ts), |
| 588 | PTRADD(ms->at, pb, ms->ts), ms->ts); |
| 589 | dest--; |
| 590 | pb--; |
| 591 | --nb; |
| 592 | if (nb == 1) |
| 593 | goto CopyA; |
| 594 | |
| 595 | k = gallop_left(PTRADD(ms->bh, pa, ms->hs), |
| 596 | PTRADD(ms->ah, baseb, ms->hs), |
| 597 | nb, nb - 1, ms); |
| 598 | k = nb - k; |
| 599 | bcount = k; |
| 600 | if (k) { |
| 601 | dest -= k; |
| 602 | pb -= k; |
| 603 | memmove(PTRADD(ms->bh, dest + 1, |
| 604 | ms->hs), |
| 605 | PTRADD(ms->ah, pb + 1, ms->hs), |
| 606 | k * ms->hs); |
| 607 | memmove(PTRADD(ms->bt, dest + 1, |
| 608 | ms->ts), |
| 609 | PTRADD(ms->at, pb + 1, ms->ts), |
| 610 | k * ms->ts); |
| 611 | nb -= k; |
| 612 | if (nb == 1) |
| 613 | goto CopyA; |
| 614 | /* nb==0 is impossible now if |
| 615 | * the comparison function is |
| 616 | * consistent, but we can't |
| 617 | * assume that it is. */ |
| 618 | if (nb == 0) |
| 619 | goto SucceedB; |
| 620 | } |
| 621 | COPY(PTRADD(ms->bh, dest, ms->hs), |
| 622 | PTRADD(ms->bh, pa, ms->hs), ms->hs); |
| 623 | COPY_any(PTRADD(ms->bt, dest, ms->ts), |
| 624 | PTRADD(ms->bt, pa, ms->ts), ms->ts); |
| 625 | dest--; |
| 626 | pa--; |
| 627 | --na; |
| 628 | if (na == 0) |
| 629 | goto SucceedB; |
| 630 | } while (acount >= MIN_GALLOP || bcount >= MIN_GALLOP); |
| 631 | ++min_gallop; /* penalize it for leaving galloping mode */ |
| 632 | ms->min_gallop = min_gallop; |
| 633 | } |
| 634 | SucceedB: |
| 635 | if (nb) { |
| 636 | COPY_anyN(PTRADD(ms->bh, dest + 1 - nb, ms->hs), |
| 637 | PTRADD(ms->ah, baseb, ms->hs), ms->hs, nb); |
| 638 | COPY_anyN(PTRADD(ms->bt, dest + 1 - nb, ms->ts), |
| 639 | PTRADD(ms->at, baseb, ms->ts), ms->ts, nb); |
| 640 | } |
| 641 | return 0; |
| 642 | CopyA: |
| 643 | assert(nb == 1 && na > 0); |
| 644 | /* The first element of pb belongs at the front of the |
| 645 | * merge. */ |
| 646 | dest -= na; |
| 647 | pa -= na; |
| 648 | memmove(PTRADD(ms->bh, dest + 1, ms->hs), |
| 649 | PTRADD(ms->bh, pa + 1, ms->hs), |
| 650 | na * ms->hs); |
| 651 | memmove(PTRADD(ms->bt, dest + 1, ms->ts), |
| 652 | PTRADD(ms->bt, pa + 1, ms->ts), |
| 653 | na * ms->ts); |
| 654 | COPY(PTRADD(ms->bh, dest, ms->hs), |
| 655 | PTRADD(ms->ah, pb, ms->hs), ms->hs); |
| 656 | COPY_any(PTRADD(ms->bt, dest, ms->ts), |
| 657 | PTRADD(ms->at, pb, ms->ts), ms->ts); |
| 658 | return 0; |
| 659 | } |
| 660 | } |
| 661 | |
| 662 | static int |
| 663 | do_ssort(MergeState *ms, ssize_t nremaining, size_t lo, size_t hi, ssize_t minrun) |
| 664 | { |
| 665 | do { |
| 666 | int descending; |
| 667 | ssize_t n; |
| 668 | |
| 669 | /* Identify next run. */ |
| 670 | { |
| 671 | /* Return the length of the run beginning at lo, in the slice [lo, |
| 672 | * hi). lo < hi is required on entry. "A run" is the longest |
| 673 | * ascending sequence, with |
| 674 | * |
| 675 | * lo[0] <= lo[1] <= lo[2] <= ... |
| 676 | * |
| 677 | * or the longest descending sequence, with |
| 678 | * |
| 679 | * lo[0] > lo[1] > lo[2] > ... |
| 680 | * |
| 681 | * Boolean descending is set to 0 in the former case, or to 1 in the |
| 682 | * latter. For its intended use in a stable mergesort, the strictness |
| 683 | * of the defn of "descending" is needed so that the caller can safely |
| 684 | * reverse a descending sequence without violating stability (strict > |
| 685 | * ensures there are no equal elements to get out of order). */ |
| 686 | size_t nlo; |
| 687 | size_t olo; |
| 688 | |
| 689 | assert(lo < hi); |
| 690 | descending = 0; |
| 691 | olo = lo; |
| 692 | nlo = lo + 1; |
| 693 | if (nlo == hi) { |
| 694 | n = 1; |
| 695 | } else { |
| 696 | n = 2; |
| 697 | if (ISLT(PTRADD(ms->bh, nlo, ms->hs), |
| 698 | PTRADD(ms->bh, olo, ms->hs), ms)) { |
| 699 | descending = 1; |
| 700 | for (olo = nlo++; |
| 701 | nlo < hi; |
| 702 | olo = nlo++, ++n) { |
| 703 | if (!ISLT(PTRADD(ms->bh, nlo, |
| 704 | ms->hs), |
| 705 | PTRADD(ms->bh, olo, |
| 706 | ms->hs), ms)) |
| 707 | break; |
| 708 | } |
| 709 | } |
| 710 | else { |
| 711 | for (olo = nlo++; |
| 712 | nlo < hi; |
| 713 | olo = nlo++, ++n) { |
| 714 | if (ISLT(PTRADD(ms->bh, nlo, |
| 715 | ms->hs), |
| 716 | PTRADD(ms->bh, olo, |
| 717 | ms->hs), ms)) |
| 718 | break; |
| 719 | } |
| 720 | } |
| 721 | } |
| 722 | } |
| 723 | if (descending) |
| 724 | reverse_slice(lo, lo + n, ms); |
| 725 | /* If short, extend to min(minrun, nremaining). */ |
| 726 | if (n < minrun) { |
| 727 | ssize_t force = nremaining <= minrun ? nremaining : minrun; |
| 728 | |
| 729 | binarysort(lo, lo + force, lo + n, ms); |
| 730 | n = force; |
| 731 | } |
| 732 | /* Push run onto pending-runs stack, and maybe merge. */ |
| 733 | assert(ms->n < MAX_MERGE_PENDING); |
| 734 | ms->pending[ms->n].base = lo; |
| 735 | ms->pending[ms->n].len = n; |
| 736 | ms->n++; |
| 737 | { |
| 738 | /* Examine the stack of runs waiting to be merged, merging adjacent |
| 739 | * runs until the stack invariants are re-established: |
| 740 | * |
| 741 | * 1. len[-3] > len[-2] + len[-1] |
| 742 | * 2. len[-2] > len[-1] |
| 743 | * |
| 744 | * See listsort.txt for more info. |
| 745 | * |
| 746 | * Returns 0 on success, -1 on error. */ |
| 747 | struct slice *p = ms->pending; |
| 748 | |
| 749 | while (ms->n > 1) { |
| 750 | ssize_t i = ms->n - 2; |
| 751 | |
| 752 | if ((i > 0 && |
| 753 | p[i-1].len <= p[i].len + p[i+1].len) || |
| 754 | (i > 1 && |
| 755 | p[i-2].len <= p[i-1].len + p[i].len)) { |
| 756 | if (p[i - 1].len < p[i + 1].len) |
| 757 | --i; |
| 758 | if (merge_at(ms, i) < 0) |
| 759 | return -1; |
| 760 | } else if (p[i].len <= p[i + 1].len) { |
| 761 | if (merge_at(ms, i) < 0) |
| 762 | return -1; |
| 763 | } else |
| 764 | break; |
| 765 | } |
| 766 | } |
| 767 | /* Advance to find next run. */ |
| 768 | lo += n; |
| 769 | nremaining -= n; |
| 770 | } while (nremaining > 0); |
| 771 | assert(lo == hi); |
| 772 | |
| 773 | { |
| 774 | /* Regardless of invariants, merge all runs on the stack until only |
| 775 | * one remains. This is used at the end of the mergesort. |
| 776 | * |
| 777 | * Returns 0 on success, -1 on error. */ |
| 778 | struct slice *p = ms->pending; |
| 779 | |
| 780 | while (ms->n > 1) { |
| 781 | ssize_t n = ms->n - 2; |
| 782 | |
| 783 | if (n > 0 && p[n - 1].len < p[n + 1].len) |
| 784 | --n; |
| 785 | if (merge_at(ms, n) < 0) |
| 786 | return -1; |
| 787 | } |
| 788 | } |
| 789 | return 0; |
| 790 | } |
| 791 | |
| 792 | #ifdef GDKssortimpl |
| 793 | /* Stable sort an array "h" (and move t accordingly). |
| 794 | * "nitems" is the number of items to sort; "hs"+"ts" is the size of |
| 795 | * the items, "tpe" is the type of the key within the items. If "heap" |
| 796 | * is non-NULL, the key is actually an offset relative to "heap" and |
| 797 | * the actual key is found at that offset (MonetDB var-sized |
| 798 | * atoms). */ |
| 799 | gdk_return |
| 800 | GDKssortimpl(void *restrict h, void *restrict t, const void *restrict heap, |
| 801 | size_t nitems, int hs, int ts, int tpe) |
| 802 | { |
| 803 | char temp; |
| 804 | MergeState ms; |
| 805 | ssize_t nremaining; |
| 806 | gdk_return result = GDK_FAIL; |
| 807 | size_t lo, hi; |
| 808 | ssize_t minrun; |
| 809 | |
| 810 | assert(h); |
| 811 | assert(hs > 0); |
| 812 | |
| 813 | ms.ah = (void *) ms.temparrayh; |
| 814 | ms.allocedh = MERGESTATE_TEMP_SIZE; |
| 815 | ms.at = (void *) ms.temparrayt; |
| 816 | ms.allocedt = MERGESTATE_TEMP_SIZE; |
| 817 | ms.n = 0; |
| 818 | ms.min_gallop = MIN_GALLOP; |
| 819 | ms.compare = ATOMcompare(tpe); |
| 820 | ms.heap = heap; |
| 821 | ms.hs = hs; |
| 822 | ms.ts = ts; |
| 823 | ms.bh = h; |
| 824 | if (!t) |
| 825 | t = &temp; |
| 826 | ms.bt = t; |
| 827 | ms.th = ms.tempstorageh; |
| 828 | ms.tt = ms.tempstoraget; |
| 829 | assert((size_t) hs <= sizeof(ms.tempstorageh)); |
| 830 | assert((size_t) ts <= sizeof(ms.tempstoraget)); |
| 831 | nremaining = (ssize_t) nitems; |
| 832 | |
| 833 | if (nremaining < 2) |
| 834 | goto succeed; |
| 835 | |
| 836 | tpe = ATOMbasetype(tpe); |
| 837 | |
| 838 | /* March over the array once, left to right, finding natural |
| 839 | * runs, and extending short natural runs to minrun |
| 840 | * elements. */ |
| 841 | lo = 0; |
| 842 | hi = lo + nremaining; |
| 843 | minrun = merge_compute_minrun(nremaining); |
| 844 | switch (tpe) { |
| 845 | case TYPE_bte: |
| 846 | if (do_ssort_bte(&ms, nremaining, lo, hi, minrun) < 0) |
| 847 | goto fail; |
| 848 | break; |
| 849 | case TYPE_sht: |
| 850 | if (do_ssort_sht(&ms, nremaining, lo, hi, minrun) < 0) |
| 851 | goto fail; |
| 852 | break; |
| 853 | case TYPE_int: |
| 854 | if (do_ssort_int(&ms, nremaining, lo, hi, minrun) < 0) |
| 855 | goto fail; |
| 856 | break; |
| 857 | case TYPE_lng: |
| 858 | if (do_ssort_lng(&ms, nremaining, lo, hi, minrun) < 0) |
| 859 | goto fail; |
| 860 | break; |
| 861 | #ifdef HAVE_HGE |
| 862 | case TYPE_hge: |
| 863 | if (do_ssort_hge(&ms, nremaining, lo, hi, minrun) < 0) |
| 864 | goto fail; |
| 865 | break; |
| 866 | #endif |
| 867 | case TYPE_flt: |
| 868 | if (do_ssort_flt(&ms, nremaining, lo, hi, minrun) < 0) |
| 869 | goto fail; |
| 870 | break; |
| 871 | case TYPE_dbl: |
| 872 | if (do_ssort_dbl(&ms, nremaining, lo, hi, minrun) < 0) |
| 873 | goto fail; |
| 874 | break; |
| 875 | default: |
| 876 | if (do_ssort_any(&ms, nremaining, lo, hi, minrun) < 0) |
| 877 | goto fail; |
| 878 | break; |
| 879 | } |
| 880 | assert(ms.n == 1); |
| 881 | assert(ms.pending[0].base == 0); |
| 882 | assert(ms.pending[0].len == (ssize_t) nitems); |
| 883 | |
| 884 | succeed: |
| 885 | result = GDK_SUCCEED; |
| 886 | fail: |
| 887 | merge_freemem(&ms); |
| 888 | return result; |
| 889 | } |
| 890 | #endif /* GDKssortimpl */ |
| 891 | |