Commit | Line | Data |
84d4ea48 |
1 | /* pp_sort.c |
2 | * |
3 | * Copyright (c) 1991-2001, Larry Wall |
4 | * |
5 | * You may distribute under the terms of either the GNU General Public |
6 | * License or the Artistic License, as specified in the README file. |
7 | * |
8 | */ |
9 | |
10 | /* |
11 | * ...they shuffled back towards the rear of the line. 'No, not at the |
12 | * rear!' the slave-driver shouted. 'Three files up. And stay there... |
13 | */ |
14 | |
15 | #include "EXTERN.h" |
16 | #define PERL_IN_PP_SORT_C |
17 | #include "perl.h" |
18 | |
19 | static I32 sortcv(pTHX_ SV *a, SV *b); |
20 | static I32 sortcv_stacked(pTHX_ SV *a, SV *b); |
21 | static I32 sortcv_xsub(pTHX_ SV *a, SV *b); |
22 | static I32 sv_ncmp(pTHX_ SV *a, SV *b); |
23 | static I32 sv_i_ncmp(pTHX_ SV *a, SV *b); |
24 | static I32 amagic_ncmp(pTHX_ SV *a, SV *b); |
25 | static I32 amagic_i_ncmp(pTHX_ SV *a, SV *b); |
26 | static I32 amagic_cmp(pTHX_ SV *a, SV *b); |
27 | static I32 amagic_cmp_locale(pTHX_ SV *a, SV *b); |
28 | |
29 | #define sv_cmp_static Perl_sv_cmp |
30 | #define sv_cmp_locale_static Perl_sv_cmp_locale |
31 | |
32 | #define SORTHINTS(hintsvp) \ |
33 | ((PL_hintgv && \ |
34 | (hintsvp = hv_fetch(GvHV(PL_hintgv), "SORT", 4, FALSE))) ? \ |
35 | (I32)SvIV(*hintsvp) : 0) |
36 | |
c53fc8a6 |
37 | #ifndef SMALLSORT |
38 | #define SMALLSORT (200) |
39 | #endif |
40 | |
84d4ea48 |
41 | /* |
42 | * The mergesort implementation is by Peter M. Mcilroy <pmcilroy@lucent.com>. |
43 | * |
44 | * The original code was written in conjunction with BSD Computer Software |
45 | * Research Group at University of California, Berkeley. |
46 | * |
47 | * See also: "Optimistic Merge Sort" (SODA '92) |
48 | * |
49 | * The integration to Perl is by John P. Linderman <jpl@research.att.com>. |
50 | * |
51 | * The code can be distributed under the same terms as Perl itself. |
52 | * |
53 | */ |
54 | |
55 | #ifdef TESTHARNESS |
56 | #include <sys/types.h> |
57 | typedef void SV; |
58 | #define pTHX_ |
59 | #define STATIC |
60 | #define New(ID,VAR,N,TYPE) VAR=(TYPE *)malloc((N)*sizeof(TYPE)) |
61 | #define Safefree(VAR) free(VAR) |
62 | typedef int (*SVCOMPARE_t) (pTHX_ SV*, SV*); |
63 | #endif /* TESTHARNESS */ |
64 | |
65 | typedef char * aptr; /* pointer for arithmetic on sizes */ |
66 | typedef SV * gptr; /* pointers in our lists */ |
67 | |
68 | /* Binary merge internal sort, with a few special mods |
69 | ** for the special perl environment it now finds itself in. |
70 | ** |
71 | ** Things that were once options have been hotwired |
72 | ** to values suitable for this use. In particular, we'll always |
73 | ** initialize looking for natural runs, we'll always produce stable |
74 | ** output, and we'll always do Peter McIlroy's binary merge. |
75 | */ |
76 | |
77 | /* Pointer types for arithmetic and storage and convenience casts */ |
78 | |
79 | #define APTR(P) ((aptr)(P)) |
80 | #define GPTP(P) ((gptr *)(P)) |
81 | #define GPPP(P) ((gptr **)(P)) |
82 | |
83 | |
84 | /* byte offset from pointer P to (larger) pointer Q */ |
85 | #define BYTEOFF(P, Q) (APTR(Q) - APTR(P)) |
86 | |
87 | #define PSIZE sizeof(gptr) |
88 | |
89 | /* If PSIZE is power of 2, make PSHIFT that power, if that helps */ |
90 | |
91 | #ifdef PSHIFT |
92 | #define PNELEM(P, Q) (BYTEOFF(P,Q) >> (PSHIFT)) |
93 | #define PNBYTE(N) ((N) << (PSHIFT)) |
94 | #define PINDEX(P, N) (GPTP(APTR(P) + PNBYTE(N))) |
95 | #else |
96 | /* Leave optimization to compiler */ |
97 | #define PNELEM(P, Q) (GPTP(Q) - GPTP(P)) |
98 | #define PNBYTE(N) ((N) * (PSIZE)) |
99 | #define PINDEX(P, N) (GPTP(P) + (N)) |
100 | #endif |
101 | |
102 | /* Pointer into other corresponding to pointer into this */ |
103 | #define POTHER(P, THIS, OTHER) GPTP(APTR(OTHER) + BYTEOFF(THIS,P)) |
104 | |
105 | #define FROMTOUPTO(src, dst, lim) do *dst++ = *src++; while(src<lim) |
106 | |
107 | |
108 | /* Runs are identified by a pointer in the auxilliary list. |
109 | ** The pointer is at the start of the list, |
110 | ** and it points to the start of the next list. |
111 | ** NEXT is used as an lvalue, too. |
112 | */ |
113 | |
114 | #define NEXT(P) (*GPPP(P)) |
115 | |
116 | |
117 | /* PTHRESH is the minimum number of pairs with the same sense to justify |
118 | ** checking for a run and extending it. Note that PTHRESH counts PAIRS, |
119 | ** not just elements, so PTHRESH == 8 means a run of 16. |
120 | */ |
121 | |
122 | #define PTHRESH (8) |
123 | |
124 | /* RTHRESH is the number of elements in a run that must compare low |
125 | ** to the low element from the opposing run before we justify |
126 | ** doing a binary rampup instead of single stepping. |
127 | ** In random input, N in a row low should only happen with |
128 | ** probability 2^(1-N), so we can risk that we are dealing |
129 | ** with orderly input without paying much when we aren't. |
130 | */ |
131 | |
132 | #define RTHRESH (6) |
133 | |
134 | |
135 | /* |
136 | ** Overview of algorithm and variables. |
137 | ** The array of elements at list1 will be organized into runs of length 2, |
138 | ** or runs of length >= 2 * PTHRESH. We only try to form long runs when |
139 | ** PTHRESH adjacent pairs compare in the same way, suggesting overall order. |
140 | ** |
141 | ** Unless otherwise specified, pair pointers address the first of two elements. |
142 | ** |
143 | ** b and b+1 are a pair that compare with sense ``sense''. |
144 | ** b is the ``bottom'' of adjacent pairs that might form a longer run. |
145 | ** |
146 | ** p2 parallels b in the list2 array, where runs are defined by |
147 | ** a pointer chain. |
148 | ** |
149 | ** t represents the ``top'' of the adjacent pairs that might extend |
150 | ** the run beginning at b. Usually, t addresses a pair |
151 | ** that compares with opposite sense from (b,b+1). |
152 | ** However, it may also address a singleton element at the end of list1, |
153 | ** or it may be equal to ``last'', the first element beyond list1. |
154 | ** |
155 | ** r addresses the Nth pair following b. If this would be beyond t, |
156 | ** we back it off to t. Only when r is less than t do we consider the |
157 | ** run long enough to consider checking. |
158 | ** |
159 | ** q addresses a pair such that the pairs at b through q already form a run. |
160 | ** Often, q will equal b, indicating we only are sure of the pair itself. |
161 | ** However, a search on the previous cycle may have revealed a longer run, |
162 | ** so q may be greater than b. |
163 | ** |
164 | ** p is used to work back from a candidate r, trying to reach q, |
165 | ** which would mean b through r would be a run. If we discover such a run, |
166 | ** we start q at r and try to push it further towards t. |
167 | ** If b through r is NOT a run, we detect the wrong order at (p-1,p). |
168 | ** In any event, after the check (if any), we have two main cases. |
169 | ** |
170 | ** 1) Short run. b <= q < p <= r <= t. |
171 | ** b through q is a run (perhaps trivial) |
172 | ** q through p are uninteresting pairs |
173 | ** p through r is a run |
174 | ** |
175 | ** 2) Long run. b < r <= q < t. |
176 | ** b through q is a run (of length >= 2 * PTHRESH) |
177 | ** |
178 | ** Note that degenerate cases are not only possible, but likely. |
179 | ** For example, if the pair following b compares with opposite sense, |
180 | ** then b == q < p == r == t. |
181 | */ |
182 | |
183 | |
184 | static void |
185 | dynprep(pTHX_ gptr *list1, gptr *list2, size_t nmemb, SVCOMPARE_t cmp) |
186 | { |
187 | int sense; |
188 | register gptr *b, *p, *q, *t, *p2; |
189 | register gptr c, *last, *r; |
190 | gptr *savep; |
191 | |
192 | b = list1; |
193 | last = PINDEX(b, nmemb); |
194 | sense = (cmp(aTHX_ *b, *(b+1)) > 0); |
195 | for (p2 = list2; b < last; ) { |
196 | /* We just started, or just reversed sense. |
197 | ** Set t at end of pairs with the prevailing sense. |
198 | */ |
199 | for (p = b+2, t = p; ++p < last; t = ++p) { |
200 | if ((cmp(aTHX_ *t, *p) > 0) != sense) break; |
201 | } |
202 | q = b; |
203 | /* Having laid out the playing field, look for long runs */ |
204 | do { |
205 | p = r = b + (2 * PTHRESH); |
206 | if (r >= t) p = r = t; /* too short to care about */ |
207 | else { |
208 | while (((cmp(aTHX_ *(p-1), *p) > 0) == sense) && |
209 | ((p -= 2) > q)); |
210 | if (p <= q) { |
211 | /* b through r is a (long) run. |
212 | ** Extend it as far as possible. |
213 | */ |
214 | p = q = r; |
215 | while (((p += 2) < t) && |
216 | ((cmp(aTHX_ *(p-1), *p) > 0) == sense)) q = p; |
217 | r = p = q + 2; /* no simple pairs, no after-run */ |
218 | } |
219 | } |
220 | if (q > b) { /* run of greater than 2 at b */ |
221 | savep = p; |
222 | p = q += 2; |
223 | /* pick up singleton, if possible */ |
224 | if ((p == t) && |
225 | ((t + 1) == last) && |
226 | ((cmp(aTHX_ *(p-1), *p) > 0) == sense)) |
227 | savep = r = p = q = last; |
228 | p2 = NEXT(p2) = p2 + (p - b); |
229 | if (sense) while (b < --p) { |
230 | c = *b; |
231 | *b++ = *p; |
232 | *p = c; |
233 | } |
234 | p = savep; |
235 | } |
236 | while (q < p) { /* simple pairs */ |
237 | p2 = NEXT(p2) = p2 + 2; |
238 | if (sense) { |
239 | c = *q++; |
240 | *(q-1) = *q; |
241 | *q++ = c; |
242 | } else q += 2; |
243 | } |
244 | if (((b = p) == t) && ((t+1) == last)) { |
245 | NEXT(p2) = p2 + 1; |
246 | b++; |
247 | } |
248 | q = r; |
249 | } while (b < t); |
250 | sense = !sense; |
251 | } |
252 | return; |
253 | } |
254 | |
255 | |
256 | /* Overview of bmerge variables: |
257 | ** |
258 | ** list1 and list2 address the main and auxiliary arrays. |
259 | ** They swap identities after each merge pass. |
260 | ** Base points to the original list1, so we can tell if |
261 | ** the pointers ended up where they belonged (or must be copied). |
262 | ** |
263 | ** When we are merging two lists, f1 and f2 are the next elements |
264 | ** on the respective lists. l1 and l2 mark the end of the lists. |
265 | ** tp2 is the current location in the merged list. |
266 | ** |
267 | ** p1 records where f1 started. |
268 | ** After the merge, a new descriptor is built there. |
269 | ** |
270 | ** p2 is a ``parallel'' pointer in (what starts as) descriptor space. |
271 | ** It is used to identify and delimit the runs. |
272 | ** |
273 | ** In the heat of determining where q, the greater of the f1/f2 elements, |
274 | ** belongs in the other list, b, t and p, represent bottom, top and probe |
275 | ** locations, respectively, in the other list. |
276 | ** They make convenient temporary pointers in other places. |
277 | */ |
278 | |
279 | STATIC void |
280 | S_mergesortsv(pTHX_ gptr *list1, size_t nmemb, SVCOMPARE_t cmp) |
281 | { |
282 | int i, run; |
283 | int sense; |
284 | register gptr *f1, *f2, *t, *b, *p, *tp2, *l1, *l2, *q; |
285 | gptr *aux, *list2, *p2, *last; |
286 | gptr *base = list1; |
287 | gptr *p1; |
c53fc8a6 |
288 | gptr small[SMALLSORT]; |
84d4ea48 |
289 | |
290 | if (nmemb <= 1) return; /* sorted trivially */ |
c53fc8a6 |
291 | if (nmemb <= SMALLSORT) list2 = small; /* use stack for aux array */ |
292 | else { New(799,list2,nmemb,gptr); } /* allocate auxilliary array */ |
84d4ea48 |
293 | aux = list2; |
294 | dynprep(aTHX_ list1, list2, nmemb, cmp); |
295 | last = PINDEX(list2, nmemb); |
296 | while (NEXT(list2) != last) { |
297 | /* More than one run remains. Do some merging to reduce runs. */ |
298 | l2 = p1 = list1; |
299 | for (tp2 = p2 = list2; p2 != last;) { |
300 | /* The new first run begins where the old second list ended. |
301 | ** Use the p2 ``parallel'' pointer to identify the end of the run. |
302 | */ |
303 | f1 = l2; |
304 | t = NEXT(p2); |
305 | f2 = l1 = POTHER(t, list2, list1); |
306 | if (t != last) t = NEXT(t); |
307 | l2 = POTHER(t, list2, list1); |
308 | p2 = t; |
309 | while (f1 < l1 && f2 < l2) { |
310 | /* If head 1 is larger than head 2, find ALL the elements |
311 | ** in list 2 strictly less than head1, write them all, |
312 | ** then head 1. Then compare the new heads, and repeat, |
313 | ** until one or both lists are exhausted. |
314 | ** |
315 | ** In all comparisons (after establishing |
316 | ** which head to merge) the item to merge |
317 | ** (at pointer q) is the first operand of |
318 | ** the comparison. When we want to know |
319 | ** if ``q is strictly less than the other'', |
320 | ** we can't just do |
321 | ** cmp(q, other) < 0 |
322 | ** because stability demands that we treat equality |
323 | ** as high when q comes from l2, and as low when |
324 | ** q was from l1. So we ask the question by doing |
325 | ** cmp(q, other) <= sense |
326 | ** and make sense == 0 when equality should look low, |
327 | ** and -1 when equality should look high. |
328 | */ |
329 | |
330 | |
331 | if (cmp(aTHX_ *f1, *f2) <= 0) { |
332 | q = f2; b = f1; t = l1; |
333 | sense = -1; |
334 | } else { |
335 | q = f1; b = f2; t = l2; |
336 | sense = 0; |
337 | } |
338 | |
339 | |
340 | /* ramp up |
341 | ** |
342 | ** Leave t at something strictly |
343 | ** greater than q (or at the end of the list), |
344 | ** and b at something strictly less than q. |
345 | */ |
346 | for (i = 1, run = 0 ;;) { |
347 | if ((p = PINDEX(b, i)) >= t) { |
348 | /* off the end */ |
349 | if (((p = PINDEX(t, -1)) > b) && |
350 | (cmp(aTHX_ *q, *p) <= sense)) |
351 | t = p; |
352 | else b = p; |
353 | break; |
354 | } else if (cmp(aTHX_ *q, *p) <= sense) { |
355 | t = p; |
356 | break; |
357 | } else b = p; |
358 | if (++run >= RTHRESH) i += i; |
359 | } |
360 | |
361 | |
362 | /* q is known to follow b and must be inserted before t. |
363 | ** Increment b, so the range of possibilities is [b,t). |
364 | ** Round binary split down, to favor early appearance. |
365 | ** Adjust b and t until q belongs just before t. |
366 | */ |
367 | |
368 | b++; |
369 | while (b < t) { |
370 | p = PINDEX(b, (PNELEM(b, t) - 1) / 2); |
371 | if (cmp(aTHX_ *q, *p) <= sense) { |
372 | t = p; |
373 | } else b = p + 1; |
374 | } |
375 | |
376 | |
377 | /* Copy all the strictly low elements */ |
378 | |
379 | if (q == f1) { |
380 | FROMTOUPTO(f2, tp2, t); |
381 | *tp2++ = *f1++; |
382 | } else { |
383 | FROMTOUPTO(f1, tp2, t); |
384 | *tp2++ = *f2++; |
385 | } |
386 | } |
387 | |
388 | |
389 | /* Run out remaining list */ |
390 | if (f1 == l1) { |
391 | if (f2 < l2) FROMTOUPTO(f2, tp2, l2); |
392 | } else FROMTOUPTO(f1, tp2, l1); |
393 | p1 = NEXT(p1) = POTHER(tp2, list2, list1); |
394 | } |
395 | t = list1; |
396 | list1 = list2; |
397 | list2 = t; |
398 | last = PINDEX(list2, nmemb); |
399 | } |
400 | if (base == list2) { |
401 | last = PINDEX(list1, nmemb); |
402 | FROMTOUPTO(list1, list2, last); |
403 | } |
c53fc8a6 |
404 | if (aux != small) Safefree(aux); /* free iff allocated */ |
84d4ea48 |
405 | return; |
406 | } |
407 | |
408 | /* |
409 | * The quicksort implementation was derived from source code contributed |
410 | * by Tom Horsley. |
411 | * |
412 | * NOTE: this code was derived from Tom Horsley's qsort replacement |
413 | * and should not be confused with the original code. |
414 | */ |
415 | |
416 | /* Copyright (C) Tom Horsley, 1997. All rights reserved. |
417 | |
418 | Permission granted to distribute under the same terms as perl which are |
419 | (briefly): |
420 | |
421 | This program is free software; you can redistribute it and/or modify |
422 | it under the terms of either: |
423 | |
424 | a) the GNU General Public License as published by the Free |
425 | Software Foundation; either version 1, or (at your option) any |
426 | later version, or |
427 | |
428 | b) the "Artistic License" which comes with this Kit. |
429 | |
430 | Details on the perl license can be found in the perl source code which |
431 | may be located via the www.perl.com web page. |
432 | |
433 | This is the most wonderfulest possible qsort I can come up with (and |
434 | still be mostly portable) My (limited) tests indicate it consistently |
435 | does about 20% fewer calls to compare than does the qsort in the Visual |
436 | C++ library, other vendors may vary. |
437 | |
438 | Some of the ideas in here can be found in "Algorithms" by Sedgewick, |
439 | others I invented myself (or more likely re-invented since they seemed |
440 | pretty obvious once I watched the algorithm operate for a while). |
441 | |
442 | Most of this code was written while watching the Marlins sweep the Giants |
443 | in the 1997 National League Playoffs - no Braves fans allowed to use this |
444 | code (just kidding :-). |
445 | |
446 | I realize that if I wanted to be true to the perl tradition, the only |
447 | comment in this file would be something like: |
448 | |
449 | ...they shuffled back towards the rear of the line. 'No, not at the |
450 | rear!' the slave-driver shouted. 'Three files up. And stay there... |
451 | |
452 | However, I really needed to violate that tradition just so I could keep |
453 | track of what happens myself, not to mention some poor fool trying to |
454 | understand this years from now :-). |
455 | */ |
456 | |
457 | /* ********************************************************** Configuration */ |
458 | |
459 | #ifndef QSORT_ORDER_GUESS |
460 | #define QSORT_ORDER_GUESS 2 /* Select doubling version of the netBSD trick */ |
461 | #endif |
462 | |
463 | /* QSORT_MAX_STACK is the largest number of partitions that can be stacked up for |
464 | future processing - a good max upper bound is log base 2 of memory size |
465 | (32 on 32 bit machines, 64 on 64 bit machines, etc). In reality can |
466 | safely be smaller than that since the program is taking up some space and |
467 | most operating systems only let you grab some subset of contiguous |
468 | memory (not to mention that you are normally sorting data larger than |
469 | 1 byte element size :-). |
470 | */ |
471 | #ifndef QSORT_MAX_STACK |
472 | #define QSORT_MAX_STACK 32 |
473 | #endif |
474 | |
475 | /* QSORT_BREAK_EVEN is the size of the largest partition we should insertion sort. |
476 | Anything bigger and we use qsort. If you make this too small, the qsort |
477 | will probably break (or become less efficient), because it doesn't expect |
478 | the middle element of a partition to be the same as the right or left - |
479 | you have been warned). |
480 | */ |
481 | #ifndef QSORT_BREAK_EVEN |
482 | #define QSORT_BREAK_EVEN 6 |
483 | #endif |
484 | |
4eb872f6 |
485 | /* QSORT_PLAY_SAFE is the size of the largest partition we're willing |
486 | to go quadratic on. We innoculate larger partitions against |
487 | quadratic behavior by shuffling them before sorting. This is not |
488 | an absolute guarantee of non-quadratic behavior, but it would take |
489 | staggeringly bad luck to pick extreme elements as the pivot |
490 | from randomized data. |
491 | */ |
492 | #ifndef QSORT_PLAY_SAFE |
493 | #define QSORT_PLAY_SAFE 255 |
494 | #endif |
495 | |
84d4ea48 |
496 | /* ************************************************************* Data Types */ |
497 | |
498 | /* hold left and right index values of a partition waiting to be sorted (the |
499 | partition includes both left and right - right is NOT one past the end or |
500 | anything like that). |
501 | */ |
502 | struct partition_stack_entry { |
503 | int left; |
504 | int right; |
505 | #ifdef QSORT_ORDER_GUESS |
506 | int qsort_break_even; |
507 | #endif |
508 | }; |
509 | |
510 | /* ******************************************************* Shorthand Macros */ |
511 | |
512 | /* Note that these macros will be used from inside the qsort function where |
513 | we happen to know that the variable 'elt_size' contains the size of an |
514 | array element and the variable 'temp' points to enough space to hold a |
515 | temp element and the variable 'array' points to the array being sorted |
516 | and 'compare' is the pointer to the compare routine. |
517 | |
518 | Also note that there are very many highly architecture specific ways |
519 | these might be sped up, but this is simply the most generally portable |
520 | code I could think of. |
521 | */ |
522 | |
523 | /* Return < 0 == 0 or > 0 as the value of elt1 is < elt2, == elt2, > elt2 |
524 | */ |
525 | #define qsort_cmp(elt1, elt2) \ |
526 | ((*compare)(aTHX_ array[elt1], array[elt2])) |
527 | |
528 | #ifdef QSORT_ORDER_GUESS |
529 | #define QSORT_NOTICE_SWAP swapped++; |
530 | #else |
531 | #define QSORT_NOTICE_SWAP |
532 | #endif |
533 | |
534 | /* swaps contents of array elements elt1, elt2. |
535 | */ |
536 | #define qsort_swap(elt1, elt2) \ |
537 | STMT_START { \ |
538 | QSORT_NOTICE_SWAP \ |
539 | temp = array[elt1]; \ |
540 | array[elt1] = array[elt2]; \ |
541 | array[elt2] = temp; \ |
542 | } STMT_END |
543 | |
544 | /* rotate contents of elt1, elt2, elt3 such that elt1 gets elt2, elt2 gets |
545 | elt3 and elt3 gets elt1. |
546 | */ |
547 | #define qsort_rotate(elt1, elt2, elt3) \ |
548 | STMT_START { \ |
549 | QSORT_NOTICE_SWAP \ |
550 | temp = array[elt1]; \ |
551 | array[elt1] = array[elt2]; \ |
552 | array[elt2] = array[elt3]; \ |
553 | array[elt3] = temp; \ |
554 | } STMT_END |
555 | |
556 | /* ************************************************************ Debug stuff */ |
557 | |
558 | #ifdef QSORT_DEBUG |
559 | |
560 | static void |
561 | break_here() |
562 | { |
563 | return; /* good place to set a breakpoint */ |
564 | } |
565 | |
566 | #define qsort_assert(t) (void)( (t) || (break_here(), 0) ) |
567 | |
568 | static void |
569 | doqsort_all_asserts( |
570 | void * array, |
571 | size_t num_elts, |
572 | size_t elt_size, |
573 | int (*compare)(const void * elt1, const void * elt2), |
574 | int pc_left, int pc_right, int u_left, int u_right) |
575 | { |
576 | int i; |
577 | |
578 | qsort_assert(pc_left <= pc_right); |
579 | qsort_assert(u_right < pc_left); |
580 | qsort_assert(pc_right < u_left); |
581 | for (i = u_right + 1; i < pc_left; ++i) { |
582 | qsort_assert(qsort_cmp(i, pc_left) < 0); |
583 | } |
584 | for (i = pc_left; i < pc_right; ++i) { |
585 | qsort_assert(qsort_cmp(i, pc_right) == 0); |
586 | } |
587 | for (i = pc_right + 1; i < u_left; ++i) { |
588 | qsort_assert(qsort_cmp(pc_right, i) < 0); |
589 | } |
590 | } |
591 | |
592 | #define qsort_all_asserts(PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT) \ |
593 | doqsort_all_asserts(array, num_elts, elt_size, compare, \ |
594 | PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT) |
595 | |
596 | #else |
597 | |
598 | #define qsort_assert(t) ((void)0) |
599 | |
600 | #define qsort_all_asserts(PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT) ((void)0) |
601 | |
602 | #endif |
603 | |
604 | /* ****************************************************************** qsort */ |
605 | |
606 | STATIC void /* the standard unstable (u) quicksort (qsort) */ |
607 | S_qsortsvu(pTHX_ SV ** array, size_t num_elts, SVCOMPARE_t compare) |
608 | { |
609 | register SV * temp; |
610 | |
611 | struct partition_stack_entry partition_stack[QSORT_MAX_STACK]; |
612 | int next_stack_entry = 0; |
613 | |
614 | int part_left; |
615 | int part_right; |
616 | #ifdef QSORT_ORDER_GUESS |
617 | int qsort_break_even; |
618 | int swapped; |
619 | #endif |
620 | |
621 | /* Make sure we actually have work to do. |
622 | */ |
623 | if (num_elts <= 1) { |
624 | return; |
625 | } |
626 | |
4eb872f6 |
627 | /* Innoculate large partitions against quadratic behavior */ |
628 | if (num_elts > QSORT_PLAY_SAFE) { |
629 | register size_t n, j; |
630 | register SV **q; |
631 | for (n = num_elts, q = array; n > 1; ) { |
632 | j = n-- * Drand01(); |
633 | temp = q[j]; |
634 | q[j] = q[n]; |
635 | q[n] = temp; |
636 | } |
637 | } |
638 | |
84d4ea48 |
639 | /* Setup the initial partition definition and fall into the sorting loop |
640 | */ |
641 | part_left = 0; |
642 | part_right = (int)(num_elts - 1); |
643 | #ifdef QSORT_ORDER_GUESS |
644 | qsort_break_even = QSORT_BREAK_EVEN; |
645 | #else |
646 | #define qsort_break_even QSORT_BREAK_EVEN |
647 | #endif |
648 | for ( ; ; ) { |
649 | if ((part_right - part_left) >= qsort_break_even) { |
650 | /* OK, this is gonna get hairy, so lets try to document all the |
651 | concepts and abbreviations and variables and what they keep |
652 | track of: |
653 | |
654 | pc: pivot chunk - the set of array elements we accumulate in the |
655 | middle of the partition, all equal in value to the original |
656 | pivot element selected. The pc is defined by: |
657 | |
658 | pc_left - the leftmost array index of the pc |
659 | pc_right - the rightmost array index of the pc |
660 | |
661 | we start with pc_left == pc_right and only one element |
662 | in the pivot chunk (but it can grow during the scan). |
663 | |
664 | u: uncompared elements - the set of elements in the partition |
665 | we have not yet compared to the pivot value. There are two |
666 | uncompared sets during the scan - one to the left of the pc |
667 | and one to the right. |
668 | |
669 | u_right - the rightmost index of the left side's uncompared set |
670 | u_left - the leftmost index of the right side's uncompared set |
671 | |
672 | The leftmost index of the left sides's uncompared set |
673 | doesn't need its own variable because it is always defined |
674 | by the leftmost edge of the whole partition (part_left). The |
675 | same goes for the rightmost edge of the right partition |
676 | (part_right). |
677 | |
678 | We know there are no uncompared elements on the left once we |
679 | get u_right < part_left and no uncompared elements on the |
680 | right once u_left > part_right. When both these conditions |
681 | are met, we have completed the scan of the partition. |
682 | |
683 | Any elements which are between the pivot chunk and the |
684 | uncompared elements should be less than the pivot value on |
685 | the left side and greater than the pivot value on the right |
686 | side (in fact, the goal of the whole algorithm is to arrange |
687 | for that to be true and make the groups of less-than and |
688 | greater-then elements into new partitions to sort again). |
689 | |
690 | As you marvel at the complexity of the code and wonder why it |
691 | has to be so confusing. Consider some of the things this level |
692 | of confusion brings: |
693 | |
694 | Once I do a compare, I squeeze every ounce of juice out of it. I |
695 | never do compare calls I don't have to do, and I certainly never |
696 | do redundant calls. |
697 | |
698 | I also never swap any elements unless I can prove there is a |
699 | good reason. Many sort algorithms will swap a known value with |
700 | an uncompared value just to get things in the right place (or |
701 | avoid complexity :-), but that uncompared value, once it gets |
702 | compared, may then have to be swapped again. A lot of the |
703 | complexity of this code is due to the fact that it never swaps |
704 | anything except compared values, and it only swaps them when the |
705 | compare shows they are out of position. |
706 | */ |
707 | int pc_left, pc_right; |
708 | int u_right, u_left; |
709 | |
710 | int s; |
711 | |
712 | pc_left = ((part_left + part_right) / 2); |
713 | pc_right = pc_left; |
714 | u_right = pc_left - 1; |
715 | u_left = pc_right + 1; |
716 | |
717 | /* Qsort works best when the pivot value is also the median value |
718 | in the partition (unfortunately you can't find the median value |
719 | without first sorting :-), so to give the algorithm a helping |
720 | hand, we pick 3 elements and sort them and use the median value |
721 | of that tiny set as the pivot value. |
722 | |
723 | Some versions of qsort like to use the left middle and right as |
724 | the 3 elements to sort so they can insure the ends of the |
725 | partition will contain values which will stop the scan in the |
726 | compare loop, but when you have to call an arbitrarily complex |
727 | routine to do a compare, its really better to just keep track of |
728 | array index values to know when you hit the edge of the |
729 | partition and avoid the extra compare. An even better reason to |
730 | avoid using a compare call is the fact that you can drop off the |
731 | edge of the array if someone foolishly provides you with an |
732 | unstable compare function that doesn't always provide consistent |
733 | results. |
734 | |
735 | So, since it is simpler for us to compare the three adjacent |
736 | elements in the middle of the partition, those are the ones we |
737 | pick here (conveniently pointed at by u_right, pc_left, and |
738 | u_left). The values of the left, center, and right elements |
739 | are refered to as l c and r in the following comments. |
740 | */ |
741 | |
742 | #ifdef QSORT_ORDER_GUESS |
743 | swapped = 0; |
744 | #endif |
745 | s = qsort_cmp(u_right, pc_left); |
746 | if (s < 0) { |
747 | /* l < c */ |
748 | s = qsort_cmp(pc_left, u_left); |
749 | /* if l < c, c < r - already in order - nothing to do */ |
750 | if (s == 0) { |
751 | /* l < c, c == r - already in order, pc grows */ |
752 | ++pc_right; |
753 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
754 | } else if (s > 0) { |
755 | /* l < c, c > r - need to know more */ |
756 | s = qsort_cmp(u_right, u_left); |
757 | if (s < 0) { |
758 | /* l < c, c > r, l < r - swap c & r to get ordered */ |
759 | qsort_swap(pc_left, u_left); |
760 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
761 | } else if (s == 0) { |
762 | /* l < c, c > r, l == r - swap c&r, grow pc */ |
763 | qsort_swap(pc_left, u_left); |
764 | --pc_left; |
765 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
766 | } else { |
767 | /* l < c, c > r, l > r - make lcr into rlc to get ordered */ |
768 | qsort_rotate(pc_left, u_right, u_left); |
769 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
770 | } |
771 | } |
772 | } else if (s == 0) { |
773 | /* l == c */ |
774 | s = qsort_cmp(pc_left, u_left); |
775 | if (s < 0) { |
776 | /* l == c, c < r - already in order, grow pc */ |
777 | --pc_left; |
778 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
779 | } else if (s == 0) { |
780 | /* l == c, c == r - already in order, grow pc both ways */ |
781 | --pc_left; |
782 | ++pc_right; |
783 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
784 | } else { |
785 | /* l == c, c > r - swap l & r, grow pc */ |
786 | qsort_swap(u_right, u_left); |
787 | ++pc_right; |
788 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
789 | } |
790 | } else { |
791 | /* l > c */ |
792 | s = qsort_cmp(pc_left, u_left); |
793 | if (s < 0) { |
794 | /* l > c, c < r - need to know more */ |
795 | s = qsort_cmp(u_right, u_left); |
796 | if (s < 0) { |
797 | /* l > c, c < r, l < r - swap l & c to get ordered */ |
798 | qsort_swap(u_right, pc_left); |
799 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
800 | } else if (s == 0) { |
801 | /* l > c, c < r, l == r - swap l & c, grow pc */ |
802 | qsort_swap(u_right, pc_left); |
803 | ++pc_right; |
804 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
805 | } else { |
806 | /* l > c, c < r, l > r - rotate lcr into crl to order */ |
807 | qsort_rotate(u_right, pc_left, u_left); |
808 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
809 | } |
810 | } else if (s == 0) { |
811 | /* l > c, c == r - swap ends, grow pc */ |
812 | qsort_swap(u_right, u_left); |
813 | --pc_left; |
814 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
815 | } else { |
816 | /* l > c, c > r - swap ends to get in order */ |
817 | qsort_swap(u_right, u_left); |
818 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
819 | } |
820 | } |
821 | /* We now know the 3 middle elements have been compared and |
822 | arranged in the desired order, so we can shrink the uncompared |
823 | sets on both sides |
824 | */ |
825 | --u_right; |
826 | ++u_left; |
827 | qsort_all_asserts(pc_left, pc_right, u_left, u_right); |
828 | |
829 | /* The above massive nested if was the simple part :-). We now have |
830 | the middle 3 elements ordered and we need to scan through the |
831 | uncompared sets on either side, swapping elements that are on |
832 | the wrong side or simply shuffling equal elements around to get |
833 | all equal elements into the pivot chunk. |
834 | */ |
835 | |
836 | for ( ; ; ) { |
837 | int still_work_on_left; |
838 | int still_work_on_right; |
839 | |
840 | /* Scan the uncompared values on the left. If I find a value |
841 | equal to the pivot value, move it over so it is adjacent to |
842 | the pivot chunk and expand the pivot chunk. If I find a value |
843 | less than the pivot value, then just leave it - its already |
844 | on the correct side of the partition. If I find a greater |
845 | value, then stop the scan. |
846 | */ |
847 | while ((still_work_on_left = (u_right >= part_left))) { |
848 | s = qsort_cmp(u_right, pc_left); |
849 | if (s < 0) { |
850 | --u_right; |
851 | } else if (s == 0) { |
852 | --pc_left; |
853 | if (pc_left != u_right) { |
854 | qsort_swap(u_right, pc_left); |
855 | } |
856 | --u_right; |
857 | } else { |
858 | break; |
859 | } |
860 | qsort_assert(u_right < pc_left); |
861 | qsort_assert(pc_left <= pc_right); |
862 | qsort_assert(qsort_cmp(u_right + 1, pc_left) <= 0); |
863 | qsort_assert(qsort_cmp(pc_left, pc_right) == 0); |
864 | } |
865 | |
866 | /* Do a mirror image scan of uncompared values on the right |
867 | */ |
868 | while ((still_work_on_right = (u_left <= part_right))) { |
869 | s = qsort_cmp(pc_right, u_left); |
870 | if (s < 0) { |
871 | ++u_left; |
872 | } else if (s == 0) { |
873 | ++pc_right; |
874 | if (pc_right != u_left) { |
875 | qsort_swap(pc_right, u_left); |
876 | } |
877 | ++u_left; |
878 | } else { |
879 | break; |
880 | } |
881 | qsort_assert(u_left > pc_right); |
882 | qsort_assert(pc_left <= pc_right); |
883 | qsort_assert(qsort_cmp(pc_right, u_left - 1) <= 0); |
884 | qsort_assert(qsort_cmp(pc_left, pc_right) == 0); |
885 | } |
886 | |
887 | if (still_work_on_left) { |
888 | /* I know I have a value on the left side which needs to be |
889 | on the right side, but I need to know more to decide |
890 | exactly the best thing to do with it. |
891 | */ |
892 | if (still_work_on_right) { |
893 | /* I know I have values on both side which are out of |
894 | position. This is a big win because I kill two birds |
895 | with one swap (so to speak). I can advance the |
896 | uncompared pointers on both sides after swapping both |
897 | of them into the right place. |
898 | */ |
899 | qsort_swap(u_right, u_left); |
900 | --u_right; |
901 | ++u_left; |
902 | qsort_all_asserts(pc_left, pc_right, u_left, u_right); |
903 | } else { |
904 | /* I have an out of position value on the left, but the |
905 | right is fully scanned, so I "slide" the pivot chunk |
906 | and any less-than values left one to make room for the |
907 | greater value over on the right. If the out of position |
908 | value is immediately adjacent to the pivot chunk (there |
909 | are no less-than values), I can do that with a swap, |
910 | otherwise, I have to rotate one of the less than values |
911 | into the former position of the out of position value |
912 | and the right end of the pivot chunk into the left end |
913 | (got all that?). |
914 | */ |
915 | --pc_left; |
916 | if (pc_left == u_right) { |
917 | qsort_swap(u_right, pc_right); |
918 | qsort_all_asserts(pc_left, pc_right-1, u_left, u_right-1); |
919 | } else { |
920 | qsort_rotate(u_right, pc_left, pc_right); |
921 | qsort_all_asserts(pc_left, pc_right-1, u_left, u_right-1); |
922 | } |
923 | --pc_right; |
924 | --u_right; |
925 | } |
926 | } else if (still_work_on_right) { |
927 | /* Mirror image of complex case above: I have an out of |
928 | position value on the right, but the left is fully |
929 | scanned, so I need to shuffle things around to make room |
930 | for the right value on the left. |
931 | */ |
932 | ++pc_right; |
933 | if (pc_right == u_left) { |
934 | qsort_swap(u_left, pc_left); |
935 | qsort_all_asserts(pc_left+1, pc_right, u_left+1, u_right); |
936 | } else { |
937 | qsort_rotate(pc_right, pc_left, u_left); |
938 | qsort_all_asserts(pc_left+1, pc_right, u_left+1, u_right); |
939 | } |
940 | ++pc_left; |
941 | ++u_left; |
942 | } else { |
943 | /* No more scanning required on either side of partition, |
944 | break out of loop and figure out next set of partitions |
945 | */ |
946 | break; |
947 | } |
948 | } |
949 | |
950 | /* The elements in the pivot chunk are now in the right place. They |
951 | will never move or be compared again. All I have to do is decide |
952 | what to do with the stuff to the left and right of the pivot |
953 | chunk. |
954 | |
955 | Notes on the QSORT_ORDER_GUESS ifdef code: |
956 | |
957 | 1. If I just built these partitions without swapping any (or |
958 | very many) elements, there is a chance that the elements are |
959 | already ordered properly (being properly ordered will |
960 | certainly result in no swapping, but the converse can't be |
961 | proved :-). |
962 | |
963 | 2. A (properly written) insertion sort will run faster on |
964 | already ordered data than qsort will. |
965 | |
966 | 3. Perhaps there is some way to make a good guess about |
967 | switching to an insertion sort earlier than partition size 6 |
968 | (for instance - we could save the partition size on the stack |
969 | and increase the size each time we find we didn't swap, thus |
970 | switching to insertion sort earlier for partitions with a |
971 | history of not swapping). |
972 | |
973 | 4. Naturally, if I just switch right away, it will make |
974 | artificial benchmarks with pure ascending (or descending) |
975 | data look really good, but is that a good reason in general? |
976 | Hard to say... |
977 | */ |
978 | |
979 | #ifdef QSORT_ORDER_GUESS |
980 | if (swapped < 3) { |
981 | #if QSORT_ORDER_GUESS == 1 |
982 | qsort_break_even = (part_right - part_left) + 1; |
983 | #endif |
984 | #if QSORT_ORDER_GUESS == 2 |
985 | qsort_break_even *= 2; |
986 | #endif |
987 | #if QSORT_ORDER_GUESS == 3 |
988 | int prev_break = qsort_break_even; |
989 | qsort_break_even *= qsort_break_even; |
990 | if (qsort_break_even < prev_break) { |
991 | qsort_break_even = (part_right - part_left) + 1; |
992 | } |
993 | #endif |
994 | } else { |
995 | qsort_break_even = QSORT_BREAK_EVEN; |
996 | } |
997 | #endif |
998 | |
999 | if (part_left < pc_left) { |
1000 | /* There are elements on the left which need more processing. |
1001 | Check the right as well before deciding what to do. |
1002 | */ |
1003 | if (pc_right < part_right) { |
1004 | /* We have two partitions to be sorted. Stack the biggest one |
1005 | and process the smallest one on the next iteration. This |
1006 | minimizes the stack height by insuring that any additional |
1007 | stack entries must come from the smallest partition which |
1008 | (because it is smallest) will have the fewest |
1009 | opportunities to generate additional stack entries. |
1010 | */ |
1011 | if ((part_right - pc_right) > (pc_left - part_left)) { |
1012 | /* stack the right partition, process the left */ |
1013 | partition_stack[next_stack_entry].left = pc_right + 1; |
1014 | partition_stack[next_stack_entry].right = part_right; |
1015 | #ifdef QSORT_ORDER_GUESS |
1016 | partition_stack[next_stack_entry].qsort_break_even = qsort_break_even; |
1017 | #endif |
1018 | part_right = pc_left - 1; |
1019 | } else { |
1020 | /* stack the left partition, process the right */ |
1021 | partition_stack[next_stack_entry].left = part_left; |
1022 | partition_stack[next_stack_entry].right = pc_left - 1; |
1023 | #ifdef QSORT_ORDER_GUESS |
1024 | partition_stack[next_stack_entry].qsort_break_even = qsort_break_even; |
1025 | #endif |
1026 | part_left = pc_right + 1; |
1027 | } |
1028 | qsort_assert(next_stack_entry < QSORT_MAX_STACK); |
1029 | ++next_stack_entry; |
1030 | } else { |
1031 | /* The elements on the left are the only remaining elements |
1032 | that need sorting, arrange for them to be processed as the |
1033 | next partition. |
1034 | */ |
1035 | part_right = pc_left - 1; |
1036 | } |
1037 | } else if (pc_right < part_right) { |
1038 | /* There is only one chunk on the right to be sorted, make it |
1039 | the new partition and loop back around. |
1040 | */ |
1041 | part_left = pc_right + 1; |
1042 | } else { |
1043 | /* This whole partition wound up in the pivot chunk, so |
1044 | we need to get a new partition off the stack. |
1045 | */ |
1046 | if (next_stack_entry == 0) { |
1047 | /* the stack is empty - we are done */ |
1048 | break; |
1049 | } |
1050 | --next_stack_entry; |
1051 | part_left = partition_stack[next_stack_entry].left; |
1052 | part_right = partition_stack[next_stack_entry].right; |
1053 | #ifdef QSORT_ORDER_GUESS |
1054 | qsort_break_even = partition_stack[next_stack_entry].qsort_break_even; |
1055 | #endif |
1056 | } |
1057 | } else { |
1058 | /* This partition is too small to fool with qsort complexity, just |
1059 | do an ordinary insertion sort to minimize overhead. |
1060 | */ |
1061 | int i; |
1062 | /* Assume 1st element is in right place already, and start checking |
1063 | at 2nd element to see where it should be inserted. |
1064 | */ |
1065 | for (i = part_left + 1; i <= part_right; ++i) { |
1066 | int j; |
1067 | /* Scan (backwards - just in case 'i' is already in right place) |
1068 | through the elements already sorted to see if the ith element |
1069 | belongs ahead of one of them. |
1070 | */ |
1071 | for (j = i - 1; j >= part_left; --j) { |
1072 | if (qsort_cmp(i, j) >= 0) { |
1073 | /* i belongs right after j |
1074 | */ |
1075 | break; |
1076 | } |
1077 | } |
1078 | ++j; |
1079 | if (j != i) { |
1080 | /* Looks like we really need to move some things |
1081 | */ |
1082 | int k; |
1083 | temp = array[i]; |
1084 | for (k = i - 1; k >= j; --k) |
1085 | array[k + 1] = array[k]; |
1086 | array[j] = temp; |
1087 | } |
1088 | } |
1089 | |
1090 | /* That partition is now sorted, grab the next one, or get out |
1091 | of the loop if there aren't any more. |
1092 | */ |
1093 | |
1094 | if (next_stack_entry == 0) { |
1095 | /* the stack is empty - we are done */ |
1096 | break; |
1097 | } |
1098 | --next_stack_entry; |
1099 | part_left = partition_stack[next_stack_entry].left; |
1100 | part_right = partition_stack[next_stack_entry].right; |
1101 | #ifdef QSORT_ORDER_GUESS |
1102 | qsort_break_even = partition_stack[next_stack_entry].qsort_break_even; |
1103 | #endif |
1104 | } |
1105 | } |
1106 | |
1107 | /* Believe it or not, the array is sorted at this point! */ |
1108 | } |
1109 | |
84d4ea48 |
1110 | /* Stabilize what is, presumably, an otherwise unstable sort method. |
1111 | * We do that by allocating (or having on hand) an array of pointers |
1112 | * that is the same size as the original array of elements to be sorted. |
1113 | * We initialize this parallel array with the addresses of the original |
1114 | * array elements. This indirection can make you crazy. |
1115 | * Some pictures can help. After initializing, we have |
1116 | * |
1117 | * indir list1 |
1118 | * +----+ +----+ |
1119 | * | | --------------> | | ------> first element to be sorted |
1120 | * +----+ +----+ |
1121 | * | | --------------> | | ------> second element to be sorted |
1122 | * +----+ +----+ |
1123 | * | | --------------> | | ------> third element to be sorted |
1124 | * +----+ +----+ |
1125 | * ... |
1126 | * +----+ +----+ |
1127 | * | | --------------> | | ------> n-1st element to be sorted |
1128 | * +----+ +----+ |
1129 | * | | --------------> | | ------> n-th element to be sorted |
1130 | * +----+ +----+ |
1131 | * |
1132 | * During the sort phase, we leave the elements of list1 where they are, |
1133 | * and sort the pointers in the indirect array in the same order determined |
1134 | * by the original comparison routine on the elements pointed to. |
1135 | * Because we don't move the elements of list1 around through |
1136 | * this phase, we can break ties on elements that compare equal |
1137 | * using their address in the list1 array, ensuring stabilty. |
1138 | * This leaves us with something looking like |
1139 | * |
1140 | * indir list1 |
1141 | * +----+ +----+ |
1142 | * | | --+ +---> | | ------> first element to be sorted |
1143 | * +----+ | | +----+ |
1144 | * | | --|-------|---> | | ------> second element to be sorted |
1145 | * +----+ | | +----+ |
1146 | * | | --|-------+ +-> | | ------> third element to be sorted |
1147 | * +----+ | | +----+ |
1148 | * ... |
1149 | * +----+ | | | | +----+ |
1150 | * | | ---|-+ | +--> | | ------> n-1st element to be sorted |
1151 | * +----+ | | +----+ |
1152 | * | | ---+ +----> | | ------> n-th element to be sorted |
1153 | * +----+ +----+ |
1154 | * |
1155 | * where the i-th element of the indirect array points to the element |
1156 | * that should be i-th in the sorted array. After the sort phase, |
1157 | * we have to put the elements of list1 into the places |
1158 | * dictated by the indirect array. |
1159 | */ |
1160 | |
1161 | static SVCOMPARE_t RealCmp; |
1162 | |
1163 | static I32 |
1164 | cmpindir(pTHX_ gptr a, gptr b) |
1165 | { |
1166 | I32 sense; |
1167 | gptr *ap = (gptr *)a; |
1168 | gptr *bp = (gptr *)b; |
1169 | |
1170 | if ((sense = RealCmp(aTHX_ *ap, *bp)) == 0) |
1171 | sense = (ap > bp) ? 1 : ((ap < bp) ? -1 : 0); |
1172 | return sense; |
1173 | } |
1174 | |
1175 | STATIC void |
1176 | S_qsortsv(pTHX_ gptr *list1, size_t nmemb, SVCOMPARE_t cmp) |
1177 | { |
1178 | SV **hintsvp; |
1179 | |
c53fc8a6 |
1180 | if (SORTHINTS(hintsvp) & HINT_SORT_STABLE) { |
84d4ea48 |
1181 | register gptr **pp, *q; |
1182 | register size_t n, j, i; |
1183 | gptr *small[SMALLSORT], **indir, tmp; |
1184 | SVCOMPARE_t savecmp; |
1185 | if (nmemb <= 1) return; /* sorted trivially */ |
4eb872f6 |
1186 | |
84d4ea48 |
1187 | /* Small arrays can use the stack, big ones must be allocated */ |
1188 | if (nmemb <= SMALLSORT) indir = small; |
1189 | else { New(1799, indir, nmemb, gptr *); } |
4eb872f6 |
1190 | |
84d4ea48 |
1191 | /* Copy pointers to original array elements into indirect array */ |
1192 | for (n = nmemb, pp = indir, q = list1; n--; ) *pp++ = q++; |
4eb872f6 |
1193 | |
84d4ea48 |
1194 | savecmp = RealCmp; /* Save current comparison routine, if any */ |
1195 | RealCmp = cmp; /* Put comparison routine where cmpindir can find it */ |
4eb872f6 |
1196 | |
84d4ea48 |
1197 | /* sort, with indirection */ |
1198 | S_qsortsvu(aTHX_ (gptr *)indir, nmemb, cmpindir); |
4eb872f6 |
1199 | |
84d4ea48 |
1200 | pp = indir; |
1201 | q = list1; |
1202 | for (n = nmemb; n--; ) { |
1203 | /* Assert A: all elements of q with index > n are already |
1204 | * in place. This is vacuosly true at the start, and we |
1205 | * put element n where it belongs below (if it wasn't |
1206 | * already where it belonged). Assert B: we only move |
1207 | * elements that aren't where they belong, |
1208 | * so, by A, we never tamper with elements above n. |
1209 | */ |
1210 | j = pp[n] - q; /* This sets j so that q[j] is |
1211 | * at pp[n]. *pp[j] belongs in |
1212 | * q[j], by construction. |
1213 | */ |
1214 | if (n != j) { /* all's well if n == j */ |
1215 | tmp = q[j]; /* save what's in q[j] */ |
1216 | do { |
1217 | q[j] = *pp[j]; /* put *pp[j] where it belongs */ |
1218 | i = pp[j] - q; /* the index in q of the element |
1219 | * just moved */ |
1220 | pp[j] = q + j; /* this is ok now */ |
1221 | } while ((j = i) != n); |
1222 | /* There are only finitely many (nmemb) addresses |
1223 | * in the pp array. |
1224 | * So we must eventually revisit an index we saw before. |
1225 | * Suppose the first revisited index is k != n. |
1226 | * An index is visited because something else belongs there. |
1227 | * If we visit k twice, then two different elements must |
1228 | * belong in the same place, which cannot be. |
1229 | * So j must get back to n, the loop terminates, |
1230 | * and we put the saved element where it belongs. |
1231 | */ |
1232 | q[n] = tmp; /* put what belongs into |
1233 | * the n-th element */ |
1234 | } |
1235 | } |
1236 | |
1237 | /* free iff allocated */ |
1238 | if (indir != small) { Safefree(indir); } |
1239 | /* restore prevailing comparison routine */ |
1240 | RealCmp = savecmp; |
c53fc8a6 |
1241 | } else { |
1242 | S_qsortsvu(aTHX_ list1, nmemb, cmp); |
84d4ea48 |
1243 | } |
1244 | } |
4eb872f6 |
1245 | |
1246 | /* |
84d4ea48 |
1247 | =for apidoc sortsv |
1248 | |
1249 | Sort an array. Here is an example: |
1250 | |
4eb872f6 |
1251 | sortsv(AvARRAY(av), av_len(av)+1, Perl_sv_cmp_locale); |
84d4ea48 |
1252 | |
1253 | =cut |
1254 | */ |
4eb872f6 |
1255 | |
84d4ea48 |
1256 | void |
1257 | Perl_sortsv(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp) |
1258 | { |
1259 | void (*sortsvp)(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp) = |
1260 | S_mergesortsv; |
1261 | SV **hintsvp; |
1262 | I32 hints; |
4eb872f6 |
1263 | |
84d4ea48 |
1264 | if ((hints = SORTHINTS(hintsvp))) { |
1265 | if (hints & HINT_SORT_QUICKSORT) |
1266 | sortsvp = S_qsortsv; |
1267 | else { |
1268 | if (hints & HINT_SORT_MERGESORT) |
1269 | sortsvp = S_mergesortsv; |
1270 | else |
1271 | sortsvp = S_mergesortsv; |
1272 | } |
1273 | } |
4eb872f6 |
1274 | |
84d4ea48 |
1275 | sortsvp(aTHX_ array, nmemb, cmp); |
1276 | } |
1277 | |
1278 | PP(pp_sort) |
1279 | { |
1280 | dSP; dMARK; dORIGMARK; |
1281 | register SV **up; |
1282 | SV **myorigmark = ORIGMARK; |
1283 | register I32 max; |
1284 | HV *stash; |
1285 | GV *gv; |
1286 | CV *cv = 0; |
1287 | I32 gimme = GIMME; |
1288 | OP* nextop = PL_op->op_next; |
1289 | I32 overloading = 0; |
1290 | bool hasargs = FALSE; |
1291 | I32 is_xsub = 0; |
1292 | |
1293 | if (gimme != G_ARRAY) { |
1294 | SP = MARK; |
1295 | RETPUSHUNDEF; |
1296 | } |
1297 | |
1298 | ENTER; |
1299 | SAVEVPTR(PL_sortcop); |
1300 | if (PL_op->op_flags & OPf_STACKED) { |
1301 | if (PL_op->op_flags & OPf_SPECIAL) { |
1302 | OP *kid = cLISTOP->op_first->op_sibling; /* pass pushmark */ |
1303 | kid = kUNOP->op_first; /* pass rv2gv */ |
1304 | kid = kUNOP->op_first; /* pass leave */ |
1305 | PL_sortcop = kid->op_next; |
1306 | stash = CopSTASH(PL_curcop); |
1307 | } |
1308 | else { |
1309 | cv = sv_2cv(*++MARK, &stash, &gv, 0); |
1310 | if (cv && SvPOK(cv)) { |
1311 | STRLEN n_a; |
1312 | char *proto = SvPV((SV*)cv, n_a); |
1313 | if (proto && strEQ(proto, "$$")) { |
1314 | hasargs = TRUE; |
1315 | } |
1316 | } |
1317 | if (!(cv && CvROOT(cv))) { |
1318 | if (cv && CvXSUB(cv)) { |
1319 | is_xsub = 1; |
1320 | } |
1321 | else if (gv) { |
1322 | SV *tmpstr = sv_newmortal(); |
1323 | gv_efullname3(tmpstr, gv, Nullch); |
1324 | DIE(aTHX_ "Undefined sort subroutine \"%s\" called", |
1325 | SvPVX(tmpstr)); |
1326 | } |
1327 | else { |
1328 | DIE(aTHX_ "Undefined subroutine in sort"); |
1329 | } |
1330 | } |
1331 | |
1332 | if (is_xsub) |
1333 | PL_sortcop = (OP*)cv; |
1334 | else { |
1335 | PL_sortcop = CvSTART(cv); |
1336 | SAVEVPTR(CvROOT(cv)->op_ppaddr); |
1337 | CvROOT(cv)->op_ppaddr = PL_ppaddr[OP_NULL]; |
1338 | |
1339 | SAVEVPTR(PL_curpad); |
1340 | PL_curpad = AvARRAY((AV*)AvARRAY(CvPADLIST(cv))[1]); |
1341 | } |
1342 | } |
1343 | } |
1344 | else { |
1345 | PL_sortcop = Nullop; |
1346 | stash = CopSTASH(PL_curcop); |
1347 | } |
1348 | |
1349 | up = myorigmark + 1; |
1350 | while (MARK < SP) { /* This may or may not shift down one here. */ |
1351 | /*SUPPRESS 560*/ |
1352 | if ((*up = *++MARK)) { /* Weed out nulls. */ |
1353 | SvTEMP_off(*up); |
1354 | if (!PL_sortcop && !SvPOK(*up)) { |
1355 | STRLEN n_a; |
1356 | if (SvAMAGIC(*up)) |
1357 | overloading = 1; |
1358 | else |
1359 | (void)sv_2pv(*up, &n_a); |
1360 | } |
1361 | up++; |
1362 | } |
1363 | } |
1364 | max = --up - myorigmark; |
1365 | if (PL_sortcop) { |
1366 | if (max > 1) { |
1367 | PERL_CONTEXT *cx; |
1368 | SV** newsp; |
1369 | bool oldcatch = CATCH_GET; |
1370 | |
1371 | SAVETMPS; |
1372 | SAVEOP(); |
1373 | |
1374 | CATCH_SET(TRUE); |
1375 | PUSHSTACKi(PERLSI_SORT); |
1376 | if (!hasargs && !is_xsub) { |
1377 | if (PL_sortstash != stash || !PL_firstgv || !PL_secondgv) { |
1378 | SAVESPTR(PL_firstgv); |
1379 | SAVESPTR(PL_secondgv); |
1380 | PL_firstgv = gv_fetchpv("a", TRUE, SVt_PV); |
1381 | PL_secondgv = gv_fetchpv("b", TRUE, SVt_PV); |
1382 | PL_sortstash = stash; |
1383 | } |
1384 | #ifdef USE_5005THREADS |
1385 | sv_lock((SV *)PL_firstgv); |
1386 | sv_lock((SV *)PL_secondgv); |
1387 | #endif |
1388 | SAVESPTR(GvSV(PL_firstgv)); |
1389 | SAVESPTR(GvSV(PL_secondgv)); |
1390 | } |
1391 | |
1392 | PUSHBLOCK(cx, CXt_NULL, PL_stack_base); |
1393 | if (!(PL_op->op_flags & OPf_SPECIAL)) { |
1394 | cx->cx_type = CXt_SUB; |
1395 | cx->blk_gimme = G_SCALAR; |
1396 | PUSHSUB(cx); |
1397 | if (!CvDEPTH(cv)) |
1398 | (void)SvREFCNT_inc(cv); /* in preparation for POPSUB */ |
1399 | } |
1400 | PL_sortcxix = cxstack_ix; |
1401 | |
1402 | if (hasargs && !is_xsub) { |
1403 | /* This is mostly copied from pp_entersub */ |
1404 | AV *av = (AV*)PL_curpad[0]; |
1405 | |
1406 | #ifndef USE_5005THREADS |
1407 | cx->blk_sub.savearray = GvAV(PL_defgv); |
1408 | GvAV(PL_defgv) = (AV*)SvREFCNT_inc(av); |
1409 | #endif /* USE_5005THREADS */ |
1410 | cx->blk_sub.oldcurpad = PL_curpad; |
1411 | cx->blk_sub.argarray = av; |
1412 | } |
1413 | sortsv((myorigmark+1), max, |
1414 | is_xsub ? sortcv_xsub : hasargs ? sortcv_stacked : sortcv); |
1415 | |
1416 | POPBLOCK(cx,PL_curpm); |
1417 | PL_stack_sp = newsp; |
1418 | POPSTACK; |
1419 | CATCH_SET(oldcatch); |
1420 | } |
1421 | } |
1422 | else { |
1423 | if (max > 1) { |
1424 | MEXTEND(SP, 20); /* Can't afford stack realloc on signal. */ |
1425 | sortsv(ORIGMARK+1, max, |
1426 | (PL_op->op_private & OPpSORT_NUMERIC) |
1427 | ? ( (PL_op->op_private & OPpSORT_INTEGER) |
1428 | ? ( overloading ? amagic_i_ncmp : sv_i_ncmp) |
1429 | : ( overloading ? amagic_ncmp : sv_ncmp)) |
1430 | : ( IN_LOCALE_RUNTIME |
1431 | ? ( overloading |
1432 | ? amagic_cmp_locale |
1433 | : sv_cmp_locale_static) |
1434 | : ( overloading ? amagic_cmp : sv_cmp_static))); |
1435 | if (PL_op->op_private & OPpSORT_REVERSE) { |
1436 | SV **p = ORIGMARK+1; |
1437 | SV **q = ORIGMARK+max; |
1438 | while (p < q) { |
1439 | SV *tmp = *p; |
1440 | *p++ = *q; |
1441 | *q-- = tmp; |
1442 | } |
1443 | } |
1444 | } |
1445 | } |
1446 | LEAVE; |
1447 | PL_stack_sp = ORIGMARK + max; |
1448 | return nextop; |
1449 | } |
1450 | |
1451 | static I32 |
1452 | sortcv(pTHX_ SV *a, SV *b) |
1453 | { |
1454 | I32 oldsaveix = PL_savestack_ix; |
1455 | I32 oldscopeix = PL_scopestack_ix; |
1456 | I32 result; |
1457 | GvSV(PL_firstgv) = a; |
1458 | GvSV(PL_secondgv) = b; |
1459 | PL_stack_sp = PL_stack_base; |
1460 | PL_op = PL_sortcop; |
1461 | CALLRUNOPS(aTHX); |
1462 | if (PL_stack_sp != PL_stack_base + 1) |
1463 | Perl_croak(aTHX_ "Sort subroutine didn't return single value"); |
1464 | if (!SvNIOKp(*PL_stack_sp)) |
1465 | Perl_croak(aTHX_ "Sort subroutine didn't return a numeric value"); |
1466 | result = SvIV(*PL_stack_sp); |
1467 | while (PL_scopestack_ix > oldscopeix) { |
1468 | LEAVE; |
1469 | } |
1470 | leave_scope(oldsaveix); |
1471 | return result; |
1472 | } |
1473 | |
1474 | static I32 |
1475 | sortcv_stacked(pTHX_ SV *a, SV *b) |
1476 | { |
1477 | I32 oldsaveix = PL_savestack_ix; |
1478 | I32 oldscopeix = PL_scopestack_ix; |
1479 | I32 result; |
1480 | AV *av; |
1481 | |
1482 | #ifdef USE_5005THREADS |
1483 | av = (AV*)PL_curpad[0]; |
1484 | #else |
1485 | av = GvAV(PL_defgv); |
1486 | #endif |
1487 | |
1488 | if (AvMAX(av) < 1) { |
1489 | SV** ary = AvALLOC(av); |
1490 | if (AvARRAY(av) != ary) { |
1491 | AvMAX(av) += AvARRAY(av) - AvALLOC(av); |
1492 | SvPVX(av) = (char*)ary; |
1493 | } |
1494 | if (AvMAX(av) < 1) { |
1495 | AvMAX(av) = 1; |
1496 | Renew(ary,2,SV*); |
1497 | SvPVX(av) = (char*)ary; |
1498 | } |
1499 | } |
1500 | AvFILLp(av) = 1; |
1501 | |
1502 | AvARRAY(av)[0] = a; |
1503 | AvARRAY(av)[1] = b; |
1504 | PL_stack_sp = PL_stack_base; |
1505 | PL_op = PL_sortcop; |
1506 | CALLRUNOPS(aTHX); |
1507 | if (PL_stack_sp != PL_stack_base + 1) |
1508 | Perl_croak(aTHX_ "Sort subroutine didn't return single value"); |
1509 | if (!SvNIOKp(*PL_stack_sp)) |
1510 | Perl_croak(aTHX_ "Sort subroutine didn't return a numeric value"); |
1511 | result = SvIV(*PL_stack_sp); |
1512 | while (PL_scopestack_ix > oldscopeix) { |
1513 | LEAVE; |
1514 | } |
1515 | leave_scope(oldsaveix); |
1516 | return result; |
1517 | } |
1518 | |
1519 | static I32 |
1520 | sortcv_xsub(pTHX_ SV *a, SV *b) |
1521 | { |
1522 | dSP; |
1523 | I32 oldsaveix = PL_savestack_ix; |
1524 | I32 oldscopeix = PL_scopestack_ix; |
1525 | I32 result; |
1526 | CV *cv=(CV*)PL_sortcop; |
1527 | |
1528 | SP = PL_stack_base; |
1529 | PUSHMARK(SP); |
1530 | EXTEND(SP, 2); |
1531 | *++SP = a; |
1532 | *++SP = b; |
1533 | PUTBACK; |
1534 | (void)(*CvXSUB(cv))(aTHX_ cv); |
1535 | if (PL_stack_sp != PL_stack_base + 1) |
1536 | Perl_croak(aTHX_ "Sort subroutine didn't return single value"); |
1537 | if (!SvNIOKp(*PL_stack_sp)) |
1538 | Perl_croak(aTHX_ "Sort subroutine didn't return a numeric value"); |
1539 | result = SvIV(*PL_stack_sp); |
1540 | while (PL_scopestack_ix > oldscopeix) { |
1541 | LEAVE; |
1542 | } |
1543 | leave_scope(oldsaveix); |
1544 | return result; |
1545 | } |
1546 | |
1547 | |
1548 | static I32 |
1549 | sv_ncmp(pTHX_ SV *a, SV *b) |
1550 | { |
1551 | NV nv1 = SvNV(a); |
1552 | NV nv2 = SvNV(b); |
1553 | return nv1 < nv2 ? -1 : nv1 > nv2 ? 1 : 0; |
1554 | } |
1555 | |
1556 | static I32 |
1557 | sv_i_ncmp(pTHX_ SV *a, SV *b) |
1558 | { |
1559 | IV iv1 = SvIV(a); |
1560 | IV iv2 = SvIV(b); |
1561 | return iv1 < iv2 ? -1 : iv1 > iv2 ? 1 : 0; |
1562 | } |
1563 | #define tryCALL_AMAGICbin(left,right,meth,svp) STMT_START { \ |
1564 | *svp = Nullsv; \ |
1565 | if (PL_amagic_generation) { \ |
1566 | if (SvAMAGIC(left)||SvAMAGIC(right))\ |
1567 | *svp = amagic_call(left, \ |
1568 | right, \ |
1569 | CAT2(meth,_amg), \ |
1570 | 0); \ |
1571 | } \ |
1572 | } STMT_END |
1573 | |
1574 | static I32 |
1575 | amagic_ncmp(pTHX_ register SV *a, register SV *b) |
1576 | { |
1577 | SV *tmpsv; |
1578 | tryCALL_AMAGICbin(a,b,ncmp,&tmpsv); |
1579 | if (tmpsv) { |
1580 | NV d; |
4eb872f6 |
1581 | |
84d4ea48 |
1582 | if (SvIOK(tmpsv)) { |
1583 | I32 i = SvIVX(tmpsv); |
1584 | if (i > 0) |
1585 | return 1; |
1586 | return i? -1 : 0; |
1587 | } |
1588 | d = SvNV(tmpsv); |
1589 | if (d > 0) |
1590 | return 1; |
1591 | return d? -1 : 0; |
1592 | } |
1593 | return sv_ncmp(aTHX_ a, b); |
1594 | } |
1595 | |
1596 | static I32 |
1597 | amagic_i_ncmp(pTHX_ register SV *a, register SV *b) |
1598 | { |
1599 | SV *tmpsv; |
1600 | tryCALL_AMAGICbin(a,b,ncmp,&tmpsv); |
1601 | if (tmpsv) { |
1602 | NV d; |
4eb872f6 |
1603 | |
84d4ea48 |
1604 | if (SvIOK(tmpsv)) { |
1605 | I32 i = SvIVX(tmpsv); |
1606 | if (i > 0) |
1607 | return 1; |
1608 | return i? -1 : 0; |
1609 | } |
1610 | d = SvNV(tmpsv); |
1611 | if (d > 0) |
1612 | return 1; |
1613 | return d? -1 : 0; |
1614 | } |
1615 | return sv_i_ncmp(aTHX_ a, b); |
1616 | } |
1617 | |
1618 | static I32 |
1619 | amagic_cmp(pTHX_ register SV *str1, register SV *str2) |
1620 | { |
1621 | SV *tmpsv; |
1622 | tryCALL_AMAGICbin(str1,str2,scmp,&tmpsv); |
1623 | if (tmpsv) { |
1624 | NV d; |
4eb872f6 |
1625 | |
84d4ea48 |
1626 | if (SvIOK(tmpsv)) { |
1627 | I32 i = SvIVX(tmpsv); |
1628 | if (i > 0) |
1629 | return 1; |
1630 | return i? -1 : 0; |
1631 | } |
1632 | d = SvNV(tmpsv); |
1633 | if (d > 0) |
1634 | return 1; |
1635 | return d? -1 : 0; |
1636 | } |
1637 | return sv_cmp(str1, str2); |
1638 | } |
1639 | |
1640 | static I32 |
1641 | amagic_cmp_locale(pTHX_ register SV *str1, register SV *str2) |
1642 | { |
1643 | SV *tmpsv; |
1644 | tryCALL_AMAGICbin(str1,str2,scmp,&tmpsv); |
1645 | if (tmpsv) { |
1646 | NV d; |
4eb872f6 |
1647 | |
84d4ea48 |
1648 | if (SvIOK(tmpsv)) { |
1649 | I32 i = SvIVX(tmpsv); |
1650 | if (i > 0) |
1651 | return 1; |
1652 | return i? -1 : 0; |
1653 | } |
1654 | d = SvNV(tmpsv); |
1655 | if (d > 0) |
1656 | return 1; |
1657 | return d? -1 : 0; |
1658 | } |
1659 | return sv_cmp_locale(str1, str2); |
1660 | } |
1661 | |
1662 | |