Commit | Line | Data |
84d4ea48 |
1 | /* pp_sort.c |
2 | * |
4bb101f2 |
3 | * Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999, |
241d1a3b |
4 | * 2000, 2001, 2002, 2003, 2004, 2005, by Larry Wall and others |
84d4ea48 |
5 | * |
6 | * You may distribute under the terms of either the GNU General Public |
7 | * License or the Artistic License, as specified in the README file. |
8 | * |
9 | */ |
10 | |
11 | /* |
12 | * ...they shuffled back towards the rear of the line. 'No, not at the |
13 | * rear!' the slave-driver shouted. 'Three files up. And stay there... |
14 | */ |
15 | |
166f8a29 |
16 | /* This file contains pp ("push/pop") functions that |
17 | * execute the opcodes that make up a perl program. A typical pp function |
18 | * expects to find its arguments on the stack, and usually pushes its |
19 | * results onto the stack, hence the 'pp' terminology. Each OP structure |
20 | * contains a pointer to the relevant pp_foo() function. |
21 | * |
22 | * This particular file just contains pp_sort(), which is complex |
23 | * enough to merit its own file! See the other pp*.c files for the rest of |
24 | * the pp_ functions. |
25 | */ |
26 | |
84d4ea48 |
27 | #include "EXTERN.h" |
28 | #define PERL_IN_PP_SORT_C |
29 | #include "perl.h" |
30 | |
42165d27 |
31 | #if defined(UNDER_CE) |
32 | /* looks like 'small' is reserved word for WINCE (or somesuch)*/ |
33 | #define small xsmall |
34 | #endif |
35 | |
84d4ea48 |
36 | static I32 sortcv(pTHX_ SV *a, SV *b); |
37 | static I32 sortcv_stacked(pTHX_ SV *a, SV *b); |
38 | static I32 sortcv_xsub(pTHX_ SV *a, SV *b); |
39 | static I32 sv_ncmp(pTHX_ SV *a, SV *b); |
40 | static I32 sv_i_ncmp(pTHX_ SV *a, SV *b); |
41 | static I32 amagic_ncmp(pTHX_ SV *a, SV *b); |
42 | static I32 amagic_i_ncmp(pTHX_ SV *a, SV *b); |
43 | static I32 amagic_cmp(pTHX_ SV *a, SV *b); |
44 | static I32 amagic_cmp_locale(pTHX_ SV *a, SV *b); |
45 | |
46 | #define sv_cmp_static Perl_sv_cmp |
47 | #define sv_cmp_locale_static Perl_sv_cmp_locale |
48 | |
045ac317 |
49 | #define SORTHINTS(hintsv) \ |
50 | (((hintsv) = GvSV(gv_fetchpv("sort::hints", GV_ADDMULTI, SVt_IV))), \ |
51 | (SvIOK(hintsv) ? ((I32)SvIV(hintsv)) : 0)) |
84d4ea48 |
52 | |
c53fc8a6 |
53 | #ifndef SMALLSORT |
54 | #define SMALLSORT (200) |
55 | #endif |
56 | |
84d4ea48 |
57 | /* |
58 | * The mergesort implementation is by Peter M. Mcilroy <pmcilroy@lucent.com>. |
59 | * |
60 | * The original code was written in conjunction with BSD Computer Software |
61 | * Research Group at University of California, Berkeley. |
62 | * |
63 | * See also: "Optimistic Merge Sort" (SODA '92) |
64 | * |
65 | * The integration to Perl is by John P. Linderman <jpl@research.att.com>. |
66 | * |
67 | * The code can be distributed under the same terms as Perl itself. |
68 | * |
69 | */ |
70 | |
84d4ea48 |
71 | |
72 | typedef char * aptr; /* pointer for arithmetic on sizes */ |
73 | typedef SV * gptr; /* pointers in our lists */ |
74 | |
75 | /* Binary merge internal sort, with a few special mods |
76 | ** for the special perl environment it now finds itself in. |
77 | ** |
78 | ** Things that were once options have been hotwired |
79 | ** to values suitable for this use. In particular, we'll always |
80 | ** initialize looking for natural runs, we'll always produce stable |
81 | ** output, and we'll always do Peter McIlroy's binary merge. |
82 | */ |
83 | |
84 | /* Pointer types for arithmetic and storage and convenience casts */ |
85 | |
86 | #define APTR(P) ((aptr)(P)) |
87 | #define GPTP(P) ((gptr *)(P)) |
88 | #define GPPP(P) ((gptr **)(P)) |
89 | |
90 | |
91 | /* byte offset from pointer P to (larger) pointer Q */ |
92 | #define BYTEOFF(P, Q) (APTR(Q) - APTR(P)) |
93 | |
94 | #define PSIZE sizeof(gptr) |
95 | |
96 | /* If PSIZE is power of 2, make PSHIFT that power, if that helps */ |
97 | |
98 | #ifdef PSHIFT |
99 | #define PNELEM(P, Q) (BYTEOFF(P,Q) >> (PSHIFT)) |
100 | #define PNBYTE(N) ((N) << (PSHIFT)) |
101 | #define PINDEX(P, N) (GPTP(APTR(P) + PNBYTE(N))) |
102 | #else |
103 | /* Leave optimization to compiler */ |
104 | #define PNELEM(P, Q) (GPTP(Q) - GPTP(P)) |
105 | #define PNBYTE(N) ((N) * (PSIZE)) |
106 | #define PINDEX(P, N) (GPTP(P) + (N)) |
107 | #endif |
108 | |
109 | /* Pointer into other corresponding to pointer into this */ |
110 | #define POTHER(P, THIS, OTHER) GPTP(APTR(OTHER) + BYTEOFF(THIS,P)) |
111 | |
112 | #define FROMTOUPTO(src, dst, lim) do *dst++ = *src++; while(src<lim) |
113 | |
114 | |
115 | /* Runs are identified by a pointer in the auxilliary list. |
116 | ** The pointer is at the start of the list, |
117 | ** and it points to the start of the next list. |
118 | ** NEXT is used as an lvalue, too. |
119 | */ |
120 | |
121 | #define NEXT(P) (*GPPP(P)) |
122 | |
123 | |
124 | /* PTHRESH is the minimum number of pairs with the same sense to justify |
125 | ** checking for a run and extending it. Note that PTHRESH counts PAIRS, |
126 | ** not just elements, so PTHRESH == 8 means a run of 16. |
127 | */ |
128 | |
129 | #define PTHRESH (8) |
130 | |
131 | /* RTHRESH is the number of elements in a run that must compare low |
132 | ** to the low element from the opposing run before we justify |
133 | ** doing a binary rampup instead of single stepping. |
134 | ** In random input, N in a row low should only happen with |
135 | ** probability 2^(1-N), so we can risk that we are dealing |
136 | ** with orderly input without paying much when we aren't. |
137 | */ |
138 | |
139 | #define RTHRESH (6) |
140 | |
141 | |
142 | /* |
143 | ** Overview of algorithm and variables. |
144 | ** The array of elements at list1 will be organized into runs of length 2, |
145 | ** or runs of length >= 2 * PTHRESH. We only try to form long runs when |
146 | ** PTHRESH adjacent pairs compare in the same way, suggesting overall order. |
147 | ** |
148 | ** Unless otherwise specified, pair pointers address the first of two elements. |
149 | ** |
150 | ** b and b+1 are a pair that compare with sense ``sense''. |
151 | ** b is the ``bottom'' of adjacent pairs that might form a longer run. |
152 | ** |
153 | ** p2 parallels b in the list2 array, where runs are defined by |
154 | ** a pointer chain. |
155 | ** |
156 | ** t represents the ``top'' of the adjacent pairs that might extend |
157 | ** the run beginning at b. Usually, t addresses a pair |
158 | ** that compares with opposite sense from (b,b+1). |
159 | ** However, it may also address a singleton element at the end of list1, |
160 | ** or it may be equal to ``last'', the first element beyond list1. |
161 | ** |
162 | ** r addresses the Nth pair following b. If this would be beyond t, |
163 | ** we back it off to t. Only when r is less than t do we consider the |
164 | ** run long enough to consider checking. |
165 | ** |
166 | ** q addresses a pair such that the pairs at b through q already form a run. |
167 | ** Often, q will equal b, indicating we only are sure of the pair itself. |
168 | ** However, a search on the previous cycle may have revealed a longer run, |
169 | ** so q may be greater than b. |
170 | ** |
171 | ** p is used to work back from a candidate r, trying to reach q, |
172 | ** which would mean b through r would be a run. If we discover such a run, |
173 | ** we start q at r and try to push it further towards t. |
174 | ** If b through r is NOT a run, we detect the wrong order at (p-1,p). |
175 | ** In any event, after the check (if any), we have two main cases. |
176 | ** |
177 | ** 1) Short run. b <= q < p <= r <= t. |
178 | ** b through q is a run (perhaps trivial) |
179 | ** q through p are uninteresting pairs |
180 | ** p through r is a run |
181 | ** |
182 | ** 2) Long run. b < r <= q < t. |
183 | ** b through q is a run (of length >= 2 * PTHRESH) |
184 | ** |
185 | ** Note that degenerate cases are not only possible, but likely. |
186 | ** For example, if the pair following b compares with opposite sense, |
187 | ** then b == q < p == r == t. |
188 | */ |
189 | |
190 | |
957d8989 |
191 | static IV |
84d4ea48 |
192 | dynprep(pTHX_ gptr *list1, gptr *list2, size_t nmemb, SVCOMPARE_t cmp) |
193 | { |
957d8989 |
194 | I32 sense; |
84d4ea48 |
195 | register gptr *b, *p, *q, *t, *p2; |
196 | register gptr c, *last, *r; |
197 | gptr *savep; |
957d8989 |
198 | IV runs = 0; |
84d4ea48 |
199 | |
200 | b = list1; |
201 | last = PINDEX(b, nmemb); |
202 | sense = (cmp(aTHX_ *b, *(b+1)) > 0); |
203 | for (p2 = list2; b < last; ) { |
204 | /* We just started, or just reversed sense. |
205 | ** Set t at end of pairs with the prevailing sense. |
206 | */ |
207 | for (p = b+2, t = p; ++p < last; t = ++p) { |
208 | if ((cmp(aTHX_ *t, *p) > 0) != sense) break; |
209 | } |
210 | q = b; |
211 | /* Having laid out the playing field, look for long runs */ |
212 | do { |
213 | p = r = b + (2 * PTHRESH); |
214 | if (r >= t) p = r = t; /* too short to care about */ |
215 | else { |
216 | while (((cmp(aTHX_ *(p-1), *p) > 0) == sense) && |
217 | ((p -= 2) > q)); |
218 | if (p <= q) { |
219 | /* b through r is a (long) run. |
220 | ** Extend it as far as possible. |
221 | */ |
222 | p = q = r; |
223 | while (((p += 2) < t) && |
224 | ((cmp(aTHX_ *(p-1), *p) > 0) == sense)) q = p; |
225 | r = p = q + 2; /* no simple pairs, no after-run */ |
226 | } |
227 | } |
228 | if (q > b) { /* run of greater than 2 at b */ |
229 | savep = p; |
230 | p = q += 2; |
231 | /* pick up singleton, if possible */ |
232 | if ((p == t) && |
233 | ((t + 1) == last) && |
234 | ((cmp(aTHX_ *(p-1), *p) > 0) == sense)) |
235 | savep = r = p = q = last; |
957d8989 |
236 | p2 = NEXT(p2) = p2 + (p - b); ++runs; |
84d4ea48 |
237 | if (sense) while (b < --p) { |
238 | c = *b; |
239 | *b++ = *p; |
240 | *p = c; |
241 | } |
242 | p = savep; |
243 | } |
244 | while (q < p) { /* simple pairs */ |
957d8989 |
245 | p2 = NEXT(p2) = p2 + 2; ++runs; |
84d4ea48 |
246 | if (sense) { |
247 | c = *q++; |
248 | *(q-1) = *q; |
249 | *q++ = c; |
250 | } else q += 2; |
251 | } |
252 | if (((b = p) == t) && ((t+1) == last)) { |
957d8989 |
253 | NEXT(p2) = p2 + 1; ++runs; |
84d4ea48 |
254 | b++; |
255 | } |
256 | q = r; |
257 | } while (b < t); |
258 | sense = !sense; |
259 | } |
957d8989 |
260 | return runs; |
84d4ea48 |
261 | } |
262 | |
263 | |
3fe0b9a9 |
264 | /* The original merge sort, in use since 5.7, was as fast as, or faster than, |
957d8989 |
265 | * qsort on many platforms, but slower than qsort, conspicuously so, |
3fe0b9a9 |
266 | * on others. The most likely explanation was platform-specific |
957d8989 |
267 | * differences in cache sizes and relative speeds. |
268 | * |
269 | * The quicksort divide-and-conquer algorithm guarantees that, as the |
270 | * problem is subdivided into smaller and smaller parts, the parts |
271 | * fit into smaller (and faster) caches. So it doesn't matter how |
272 | * many levels of cache exist, quicksort will "find" them, and, |
273 | * as long as smaller is faster, take advanatge of them. |
274 | * |
3fe0b9a9 |
275 | * By contrast, consider how the original mergesort algorithm worked. |
957d8989 |
276 | * Suppose we have five runs (each typically of length 2 after dynprep). |
277 | * |
278 | * pass base aux |
279 | * 0 1 2 3 4 5 |
280 | * 1 12 34 5 |
281 | * 2 1234 5 |
282 | * 3 12345 |
283 | * 4 12345 |
284 | * |
285 | * Adjacent pairs are merged in "grand sweeps" through the input. |
286 | * This means, on pass 1, the records in runs 1 and 2 aren't revisited until |
287 | * runs 3 and 4 are merged and the runs from run 5 have been copied. |
288 | * The only cache that matters is one large enough to hold *all* the input. |
289 | * On some platforms, this may be many times slower than smaller caches. |
290 | * |
291 | * The following pseudo-code uses the same basic merge algorithm, |
292 | * but in a divide-and-conquer way. |
293 | * |
294 | * # merge $runs runs at offset $offset of list $list1 into $list2. |
295 | * # all unmerged runs ($runs == 1) originate in list $base. |
296 | * sub mgsort2 { |
297 | * my ($offset, $runs, $base, $list1, $list2) = @_; |
298 | * |
299 | * if ($runs == 1) { |
300 | * if ($list1 is $base) copy run to $list2 |
301 | * return offset of end of list (or copy) |
302 | * } else { |
303 | * $off2 = mgsort2($offset, $runs-($runs/2), $base, $list2, $list1) |
304 | * mgsort2($off2, $runs/2, $base, $list2, $list1) |
305 | * merge the adjacent runs at $offset of $list1 into $list2 |
306 | * return the offset of the end of the merged runs |
307 | * } |
308 | * } |
309 | * mgsort2(0, $runs, $base, $aux, $base); |
310 | * |
311 | * For our 5 runs, the tree of calls looks like |
312 | * |
313 | * 5 |
314 | * 3 2 |
315 | * 2 1 1 1 |
316 | * 1 1 |
317 | * |
318 | * 1 2 3 4 5 |
319 | * |
320 | * and the corresponding activity looks like |
321 | * |
322 | * copy runs 1 and 2 from base to aux |
323 | * merge runs 1 and 2 from aux to base |
324 | * (run 3 is where it belongs, no copy needed) |
325 | * merge runs 12 and 3 from base to aux |
326 | * (runs 4 and 5 are where they belong, no copy needed) |
327 | * merge runs 4 and 5 from base to aux |
328 | * merge runs 123 and 45 from aux to base |
329 | * |
330 | * Note that we merge runs 1 and 2 immediately after copying them, |
331 | * while they are still likely to be in fast cache. Similarly, |
332 | * run 3 is merged with run 12 while it still may be lingering in cache. |
333 | * This implementation should therefore enjoy much of the cache-friendly |
334 | * behavior that quicksort does. In addition, it does less copying |
335 | * than the original mergesort implementation (only runs 1 and 2 are copied) |
336 | * and the "balancing" of merges is better (merged runs comprise more nearly |
337 | * equal numbers of original runs). |
338 | * |
339 | * The actual cache-friendly implementation will use a pseudo-stack |
340 | * to avoid recursion, and will unroll processing of runs of length 2, |
341 | * but it is otherwise similar to the recursive implementation. |
957d8989 |
342 | */ |
343 | |
344 | typedef struct { |
345 | IV offset; /* offset of 1st of 2 runs at this level */ |
346 | IV runs; /* how many runs must be combined into 1 */ |
347 | } off_runs; /* pseudo-stack element */ |
348 | |
6c3fb703 |
349 | |
350 | static I32 |
351 | cmp_desc(pTHX_ gptr a, gptr b) |
352 | { |
353 | return -PL_sort_RealCmp(aTHX_ a, b); |
354 | } |
355 | |
957d8989 |
356 | STATIC void |
6c3fb703 |
357 | S_mergesortsv(pTHX_ gptr *base, size_t nmemb, SVCOMPARE_t cmp, U32 flags) |
957d8989 |
358 | { |
359 | IV i, run, runs, offset; |
360 | I32 sense, level; |
361 | int iwhich; |
362 | register gptr *f1, *f2, *t, *b, *p, *tp2, *l1, *l2, *q; |
363 | gptr *aux, *list1, *list2; |
364 | gptr *p1; |
365 | gptr small[SMALLSORT]; |
366 | gptr *which[3]; |
367 | off_runs stack[60], *stackp; |
a80036c6 |
368 | SVCOMPARE_t savecmp = 0; |
957d8989 |
369 | |
370 | if (nmemb <= 1) return; /* sorted trivially */ |
6c3fb703 |
371 | |
372 | if (flags) { |
373 | savecmp = PL_sort_RealCmp; /* Save current comparison routine, if any */ |
374 | PL_sort_RealCmp = cmp; /* Put comparison routine where cmp_desc can find it */ |
375 | cmp = cmp_desc; |
376 | } |
377 | |
957d8989 |
378 | if (nmemb <= SMALLSORT) aux = small; /* use stack for aux array */ |
379 | else { New(799,aux,nmemb,gptr); } /* allocate auxilliary array */ |
380 | level = 0; |
381 | stackp = stack; |
382 | stackp->runs = dynprep(aTHX_ base, aux, nmemb, cmp); |
383 | stackp->offset = offset = 0; |
384 | which[0] = which[2] = base; |
385 | which[1] = aux; |
386 | for (;;) { |
387 | /* On levels where both runs have be constructed (stackp->runs == 0), |
388 | * merge them, and note the offset of their end, in case the offset |
389 | * is needed at the next level up. Hop up a level, and, |
390 | * as long as stackp->runs is 0, keep merging. |
391 | */ |
392 | if ((runs = stackp->runs) == 0) { |
393 | iwhich = level & 1; |
394 | list1 = which[iwhich]; /* area where runs are now */ |
395 | list2 = which[++iwhich]; /* area for merged runs */ |
396 | do { |
397 | offset = stackp->offset; |
398 | f1 = p1 = list1 + offset; /* start of first run */ |
399 | p = tp2 = list2 + offset; /* where merged run will go */ |
400 | t = NEXT(p); /* where first run ends */ |
401 | f2 = l1 = POTHER(t, list2, list1); /* ... on the other side */ |
402 | t = NEXT(t); /* where second runs ends */ |
403 | l2 = POTHER(t, list2, list1); /* ... on the other side */ |
404 | offset = PNELEM(list2, t); |
405 | while (f1 < l1 && f2 < l2) { |
406 | /* If head 1 is larger than head 2, find ALL the elements |
407 | ** in list 2 strictly less than head1, write them all, |
408 | ** then head 1. Then compare the new heads, and repeat, |
409 | ** until one or both lists are exhausted. |
410 | ** |
411 | ** In all comparisons (after establishing |
412 | ** which head to merge) the item to merge |
413 | ** (at pointer q) is the first operand of |
414 | ** the comparison. When we want to know |
415 | ** if ``q is strictly less than the other'', |
416 | ** we can't just do |
417 | ** cmp(q, other) < 0 |
418 | ** because stability demands that we treat equality |
419 | ** as high when q comes from l2, and as low when |
420 | ** q was from l1. So we ask the question by doing |
421 | ** cmp(q, other) <= sense |
422 | ** and make sense == 0 when equality should look low, |
423 | ** and -1 when equality should look high. |
424 | */ |
425 | |
426 | |
427 | if (cmp(aTHX_ *f1, *f2) <= 0) { |
428 | q = f2; b = f1; t = l1; |
429 | sense = -1; |
430 | } else { |
431 | q = f1; b = f2; t = l2; |
432 | sense = 0; |
433 | } |
434 | |
435 | |
436 | /* ramp up |
437 | ** |
438 | ** Leave t at something strictly |
439 | ** greater than q (or at the end of the list), |
440 | ** and b at something strictly less than q. |
441 | */ |
442 | for (i = 1, run = 0 ;;) { |
443 | if ((p = PINDEX(b, i)) >= t) { |
444 | /* off the end */ |
445 | if (((p = PINDEX(t, -1)) > b) && |
446 | (cmp(aTHX_ *q, *p) <= sense)) |
447 | t = p; |
448 | else b = p; |
449 | break; |
450 | } else if (cmp(aTHX_ *q, *p) <= sense) { |
451 | t = p; |
452 | break; |
453 | } else b = p; |
454 | if (++run >= RTHRESH) i += i; |
455 | } |
456 | |
457 | |
458 | /* q is known to follow b and must be inserted before t. |
459 | ** Increment b, so the range of possibilities is [b,t). |
460 | ** Round binary split down, to favor early appearance. |
461 | ** Adjust b and t until q belongs just before t. |
462 | */ |
463 | |
464 | b++; |
465 | while (b < t) { |
466 | p = PINDEX(b, (PNELEM(b, t) - 1) / 2); |
467 | if (cmp(aTHX_ *q, *p) <= sense) { |
468 | t = p; |
469 | } else b = p + 1; |
470 | } |
471 | |
472 | |
473 | /* Copy all the strictly low elements */ |
474 | |
475 | if (q == f1) { |
476 | FROMTOUPTO(f2, tp2, t); |
477 | *tp2++ = *f1++; |
478 | } else { |
479 | FROMTOUPTO(f1, tp2, t); |
480 | *tp2++ = *f2++; |
481 | } |
482 | } |
483 | |
484 | |
485 | /* Run out remaining list */ |
486 | if (f1 == l1) { |
487 | if (f2 < l2) FROMTOUPTO(f2, tp2, l2); |
488 | } else FROMTOUPTO(f1, tp2, l1); |
489 | p1 = NEXT(p1) = POTHER(tp2, list2, list1); |
490 | |
491 | if (--level == 0) goto done; |
492 | --stackp; |
493 | t = list1; list1 = list2; list2 = t; /* swap lists */ |
494 | } while ((runs = stackp->runs) == 0); |
495 | } |
496 | |
497 | |
498 | stackp->runs = 0; /* current run will finish level */ |
499 | /* While there are more than 2 runs remaining, |
500 | * turn them into exactly 2 runs (at the "other" level), |
501 | * each made up of approximately half the runs. |
502 | * Stack the second half for later processing, |
503 | * and set about producing the first half now. |
504 | */ |
505 | while (runs > 2) { |
506 | ++level; |
507 | ++stackp; |
508 | stackp->offset = offset; |
509 | runs -= stackp->runs = runs / 2; |
510 | } |
511 | /* We must construct a single run from 1 or 2 runs. |
512 | * All the original runs are in which[0] == base. |
513 | * The run we construct must end up in which[level&1]. |
514 | */ |
515 | iwhich = level & 1; |
516 | if (runs == 1) { |
517 | /* Constructing a single run from a single run. |
518 | * If it's where it belongs already, there's nothing to do. |
519 | * Otherwise, copy it to where it belongs. |
520 | * A run of 1 is either a singleton at level 0, |
521 | * or the second half of a split 3. In neither event |
522 | * is it necessary to set offset. It will be set by the merge |
523 | * that immediately follows. |
524 | */ |
525 | if (iwhich) { /* Belongs in aux, currently in base */ |
526 | f1 = b = PINDEX(base, offset); /* where list starts */ |
527 | f2 = PINDEX(aux, offset); /* where list goes */ |
528 | t = NEXT(f2); /* where list will end */ |
529 | offset = PNELEM(aux, t); /* offset thereof */ |
530 | t = PINDEX(base, offset); /* where it currently ends */ |
531 | FROMTOUPTO(f1, f2, t); /* copy */ |
532 | NEXT(b) = t; /* set up parallel pointer */ |
533 | } else if (level == 0) goto done; /* single run at level 0 */ |
534 | } else { |
535 | /* Constructing a single run from two runs. |
536 | * The merge code at the top will do that. |
537 | * We need only make sure the two runs are in the "other" array, |
538 | * so they'll end up in the correct array after the merge. |
539 | */ |
540 | ++level; |
541 | ++stackp; |
542 | stackp->offset = offset; |
543 | stackp->runs = 0; /* take care of both runs, trigger merge */ |
544 | if (!iwhich) { /* Merged runs belong in aux, copy 1st */ |
545 | f1 = b = PINDEX(base, offset); /* where first run starts */ |
546 | f2 = PINDEX(aux, offset); /* where it will be copied */ |
547 | t = NEXT(f2); /* where first run will end */ |
548 | offset = PNELEM(aux, t); /* offset thereof */ |
549 | p = PINDEX(base, offset); /* end of first run */ |
550 | t = NEXT(t); /* where second run will end */ |
551 | t = PINDEX(base, PNELEM(aux, t)); /* where it now ends */ |
552 | FROMTOUPTO(f1, f2, t); /* copy both runs */ |
553 | NEXT(b) = p; /* paralled pointer for 1st */ |
554 | NEXT(p) = t; /* ... and for second */ |
555 | } |
556 | } |
557 | } |
558 | done: |
559 | if (aux != small) Safefree(aux); /* free iff allocated */ |
6c3fb703 |
560 | if (flags) { |
561 | PL_sort_RealCmp = savecmp; /* Restore current comparison routine, if any */ |
562 | } |
957d8989 |
563 | return; |
564 | } |
565 | |
84d4ea48 |
566 | /* |
567 | * The quicksort implementation was derived from source code contributed |
568 | * by Tom Horsley. |
569 | * |
570 | * NOTE: this code was derived from Tom Horsley's qsort replacement |
571 | * and should not be confused with the original code. |
572 | */ |
573 | |
574 | /* Copyright (C) Tom Horsley, 1997. All rights reserved. |
575 | |
576 | Permission granted to distribute under the same terms as perl which are |
577 | (briefly): |
578 | |
579 | This program is free software; you can redistribute it and/or modify |
580 | it under the terms of either: |
581 | |
582 | a) the GNU General Public License as published by the Free |
583 | Software Foundation; either version 1, or (at your option) any |
584 | later version, or |
585 | |
586 | b) the "Artistic License" which comes with this Kit. |
587 | |
588 | Details on the perl license can be found in the perl source code which |
589 | may be located via the www.perl.com web page. |
590 | |
591 | This is the most wonderfulest possible qsort I can come up with (and |
592 | still be mostly portable) My (limited) tests indicate it consistently |
593 | does about 20% fewer calls to compare than does the qsort in the Visual |
594 | C++ library, other vendors may vary. |
595 | |
596 | Some of the ideas in here can be found in "Algorithms" by Sedgewick, |
597 | others I invented myself (or more likely re-invented since they seemed |
598 | pretty obvious once I watched the algorithm operate for a while). |
599 | |
600 | Most of this code was written while watching the Marlins sweep the Giants |
601 | in the 1997 National League Playoffs - no Braves fans allowed to use this |
602 | code (just kidding :-). |
603 | |
604 | I realize that if I wanted to be true to the perl tradition, the only |
605 | comment in this file would be something like: |
606 | |
607 | ...they shuffled back towards the rear of the line. 'No, not at the |
608 | rear!' the slave-driver shouted. 'Three files up. And stay there... |
609 | |
610 | However, I really needed to violate that tradition just so I could keep |
611 | track of what happens myself, not to mention some poor fool trying to |
612 | understand this years from now :-). |
613 | */ |
614 | |
615 | /* ********************************************************** Configuration */ |
616 | |
617 | #ifndef QSORT_ORDER_GUESS |
618 | #define QSORT_ORDER_GUESS 2 /* Select doubling version of the netBSD trick */ |
619 | #endif |
620 | |
621 | /* QSORT_MAX_STACK is the largest number of partitions that can be stacked up for |
622 | future processing - a good max upper bound is log base 2 of memory size |
623 | (32 on 32 bit machines, 64 on 64 bit machines, etc). In reality can |
624 | safely be smaller than that since the program is taking up some space and |
625 | most operating systems only let you grab some subset of contiguous |
626 | memory (not to mention that you are normally sorting data larger than |
627 | 1 byte element size :-). |
628 | */ |
629 | #ifndef QSORT_MAX_STACK |
630 | #define QSORT_MAX_STACK 32 |
631 | #endif |
632 | |
633 | /* QSORT_BREAK_EVEN is the size of the largest partition we should insertion sort. |
634 | Anything bigger and we use qsort. If you make this too small, the qsort |
635 | will probably break (or become less efficient), because it doesn't expect |
636 | the middle element of a partition to be the same as the right or left - |
637 | you have been warned). |
638 | */ |
639 | #ifndef QSORT_BREAK_EVEN |
640 | #define QSORT_BREAK_EVEN 6 |
641 | #endif |
642 | |
4eb872f6 |
643 | /* QSORT_PLAY_SAFE is the size of the largest partition we're willing |
644 | to go quadratic on. We innoculate larger partitions against |
645 | quadratic behavior by shuffling them before sorting. This is not |
646 | an absolute guarantee of non-quadratic behavior, but it would take |
647 | staggeringly bad luck to pick extreme elements as the pivot |
648 | from randomized data. |
649 | */ |
650 | #ifndef QSORT_PLAY_SAFE |
651 | #define QSORT_PLAY_SAFE 255 |
652 | #endif |
653 | |
84d4ea48 |
654 | /* ************************************************************* Data Types */ |
655 | |
656 | /* hold left and right index values of a partition waiting to be sorted (the |
657 | partition includes both left and right - right is NOT one past the end or |
658 | anything like that). |
659 | */ |
660 | struct partition_stack_entry { |
661 | int left; |
662 | int right; |
663 | #ifdef QSORT_ORDER_GUESS |
664 | int qsort_break_even; |
665 | #endif |
666 | }; |
667 | |
668 | /* ******************************************************* Shorthand Macros */ |
669 | |
670 | /* Note that these macros will be used from inside the qsort function where |
671 | we happen to know that the variable 'elt_size' contains the size of an |
672 | array element and the variable 'temp' points to enough space to hold a |
673 | temp element and the variable 'array' points to the array being sorted |
674 | and 'compare' is the pointer to the compare routine. |
675 | |
676 | Also note that there are very many highly architecture specific ways |
677 | these might be sped up, but this is simply the most generally portable |
678 | code I could think of. |
679 | */ |
680 | |
681 | /* Return < 0 == 0 or > 0 as the value of elt1 is < elt2, == elt2, > elt2 |
682 | */ |
683 | #define qsort_cmp(elt1, elt2) \ |
684 | ((*compare)(aTHX_ array[elt1], array[elt2])) |
685 | |
686 | #ifdef QSORT_ORDER_GUESS |
687 | #define QSORT_NOTICE_SWAP swapped++; |
688 | #else |
689 | #define QSORT_NOTICE_SWAP |
690 | #endif |
691 | |
692 | /* swaps contents of array elements elt1, elt2. |
693 | */ |
694 | #define qsort_swap(elt1, elt2) \ |
695 | STMT_START { \ |
696 | QSORT_NOTICE_SWAP \ |
697 | temp = array[elt1]; \ |
698 | array[elt1] = array[elt2]; \ |
699 | array[elt2] = temp; \ |
700 | } STMT_END |
701 | |
702 | /* rotate contents of elt1, elt2, elt3 such that elt1 gets elt2, elt2 gets |
703 | elt3 and elt3 gets elt1. |
704 | */ |
705 | #define qsort_rotate(elt1, elt2, elt3) \ |
706 | STMT_START { \ |
707 | QSORT_NOTICE_SWAP \ |
708 | temp = array[elt1]; \ |
709 | array[elt1] = array[elt2]; \ |
710 | array[elt2] = array[elt3]; \ |
711 | array[elt3] = temp; \ |
712 | } STMT_END |
713 | |
714 | /* ************************************************************ Debug stuff */ |
715 | |
716 | #ifdef QSORT_DEBUG |
717 | |
718 | static void |
719 | break_here() |
720 | { |
721 | return; /* good place to set a breakpoint */ |
722 | } |
723 | |
724 | #define qsort_assert(t) (void)( (t) || (break_here(), 0) ) |
725 | |
726 | static void |
727 | doqsort_all_asserts( |
728 | void * array, |
729 | size_t num_elts, |
730 | size_t elt_size, |
731 | int (*compare)(const void * elt1, const void * elt2), |
732 | int pc_left, int pc_right, int u_left, int u_right) |
733 | { |
734 | int i; |
735 | |
736 | qsort_assert(pc_left <= pc_right); |
737 | qsort_assert(u_right < pc_left); |
738 | qsort_assert(pc_right < u_left); |
739 | for (i = u_right + 1; i < pc_left; ++i) { |
740 | qsort_assert(qsort_cmp(i, pc_left) < 0); |
741 | } |
742 | for (i = pc_left; i < pc_right; ++i) { |
743 | qsort_assert(qsort_cmp(i, pc_right) == 0); |
744 | } |
745 | for (i = pc_right + 1; i < u_left; ++i) { |
746 | qsort_assert(qsort_cmp(pc_right, i) < 0); |
747 | } |
748 | } |
749 | |
750 | #define qsort_all_asserts(PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT) \ |
751 | doqsort_all_asserts(array, num_elts, elt_size, compare, \ |
752 | PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT) |
753 | |
754 | #else |
755 | |
756 | #define qsort_assert(t) ((void)0) |
757 | |
758 | #define qsort_all_asserts(PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT) ((void)0) |
759 | |
760 | #endif |
761 | |
762 | /* ****************************************************************** qsort */ |
763 | |
764 | STATIC void /* the standard unstable (u) quicksort (qsort) */ |
765 | S_qsortsvu(pTHX_ SV ** array, size_t num_elts, SVCOMPARE_t compare) |
766 | { |
767 | register SV * temp; |
768 | |
769 | struct partition_stack_entry partition_stack[QSORT_MAX_STACK]; |
770 | int next_stack_entry = 0; |
771 | |
772 | int part_left; |
773 | int part_right; |
774 | #ifdef QSORT_ORDER_GUESS |
775 | int qsort_break_even; |
776 | int swapped; |
777 | #endif |
778 | |
779 | /* Make sure we actually have work to do. |
780 | */ |
781 | if (num_elts <= 1) { |
782 | return; |
783 | } |
784 | |
4eb872f6 |
785 | /* Innoculate large partitions against quadratic behavior */ |
786 | if (num_elts > QSORT_PLAY_SAFE) { |
787 | register size_t n, j; |
788 | register SV **q; |
789 | for (n = num_elts, q = array; n > 1; ) { |
eb160463 |
790 | j = (size_t)(n-- * Drand01()); |
4eb872f6 |
791 | temp = q[j]; |
792 | q[j] = q[n]; |
793 | q[n] = temp; |
794 | } |
795 | } |
796 | |
84d4ea48 |
797 | /* Setup the initial partition definition and fall into the sorting loop |
798 | */ |
799 | part_left = 0; |
800 | part_right = (int)(num_elts - 1); |
801 | #ifdef QSORT_ORDER_GUESS |
802 | qsort_break_even = QSORT_BREAK_EVEN; |
803 | #else |
804 | #define qsort_break_even QSORT_BREAK_EVEN |
805 | #endif |
806 | for ( ; ; ) { |
807 | if ((part_right - part_left) >= qsort_break_even) { |
808 | /* OK, this is gonna get hairy, so lets try to document all the |
809 | concepts and abbreviations and variables and what they keep |
810 | track of: |
811 | |
812 | pc: pivot chunk - the set of array elements we accumulate in the |
813 | middle of the partition, all equal in value to the original |
814 | pivot element selected. The pc is defined by: |
815 | |
816 | pc_left - the leftmost array index of the pc |
817 | pc_right - the rightmost array index of the pc |
818 | |
819 | we start with pc_left == pc_right and only one element |
820 | in the pivot chunk (but it can grow during the scan). |
821 | |
822 | u: uncompared elements - the set of elements in the partition |
823 | we have not yet compared to the pivot value. There are two |
824 | uncompared sets during the scan - one to the left of the pc |
825 | and one to the right. |
826 | |
827 | u_right - the rightmost index of the left side's uncompared set |
828 | u_left - the leftmost index of the right side's uncompared set |
829 | |
830 | The leftmost index of the left sides's uncompared set |
831 | doesn't need its own variable because it is always defined |
832 | by the leftmost edge of the whole partition (part_left). The |
833 | same goes for the rightmost edge of the right partition |
834 | (part_right). |
835 | |
836 | We know there are no uncompared elements on the left once we |
837 | get u_right < part_left and no uncompared elements on the |
838 | right once u_left > part_right. When both these conditions |
839 | are met, we have completed the scan of the partition. |
840 | |
841 | Any elements which are between the pivot chunk and the |
842 | uncompared elements should be less than the pivot value on |
843 | the left side and greater than the pivot value on the right |
844 | side (in fact, the goal of the whole algorithm is to arrange |
845 | for that to be true and make the groups of less-than and |
846 | greater-then elements into new partitions to sort again). |
847 | |
848 | As you marvel at the complexity of the code and wonder why it |
849 | has to be so confusing. Consider some of the things this level |
850 | of confusion brings: |
851 | |
852 | Once I do a compare, I squeeze every ounce of juice out of it. I |
853 | never do compare calls I don't have to do, and I certainly never |
854 | do redundant calls. |
855 | |
856 | I also never swap any elements unless I can prove there is a |
857 | good reason. Many sort algorithms will swap a known value with |
858 | an uncompared value just to get things in the right place (or |
859 | avoid complexity :-), but that uncompared value, once it gets |
860 | compared, may then have to be swapped again. A lot of the |
861 | complexity of this code is due to the fact that it never swaps |
862 | anything except compared values, and it only swaps them when the |
863 | compare shows they are out of position. |
864 | */ |
865 | int pc_left, pc_right; |
866 | int u_right, u_left; |
867 | |
868 | int s; |
869 | |
870 | pc_left = ((part_left + part_right) / 2); |
871 | pc_right = pc_left; |
872 | u_right = pc_left - 1; |
873 | u_left = pc_right + 1; |
874 | |
875 | /* Qsort works best when the pivot value is also the median value |
876 | in the partition (unfortunately you can't find the median value |
877 | without first sorting :-), so to give the algorithm a helping |
878 | hand, we pick 3 elements and sort them and use the median value |
879 | of that tiny set as the pivot value. |
880 | |
881 | Some versions of qsort like to use the left middle and right as |
882 | the 3 elements to sort so they can insure the ends of the |
883 | partition will contain values which will stop the scan in the |
884 | compare loop, but when you have to call an arbitrarily complex |
885 | routine to do a compare, its really better to just keep track of |
886 | array index values to know when you hit the edge of the |
887 | partition and avoid the extra compare. An even better reason to |
888 | avoid using a compare call is the fact that you can drop off the |
889 | edge of the array if someone foolishly provides you with an |
890 | unstable compare function that doesn't always provide consistent |
891 | results. |
892 | |
893 | So, since it is simpler for us to compare the three adjacent |
894 | elements in the middle of the partition, those are the ones we |
895 | pick here (conveniently pointed at by u_right, pc_left, and |
896 | u_left). The values of the left, center, and right elements |
897 | are refered to as l c and r in the following comments. |
898 | */ |
899 | |
900 | #ifdef QSORT_ORDER_GUESS |
901 | swapped = 0; |
902 | #endif |
903 | s = qsort_cmp(u_right, pc_left); |
904 | if (s < 0) { |
905 | /* l < c */ |
906 | s = qsort_cmp(pc_left, u_left); |
907 | /* if l < c, c < r - already in order - nothing to do */ |
908 | if (s == 0) { |
909 | /* l < c, c == r - already in order, pc grows */ |
910 | ++pc_right; |
911 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
912 | } else if (s > 0) { |
913 | /* l < c, c > r - need to know more */ |
914 | s = qsort_cmp(u_right, u_left); |
915 | if (s < 0) { |
916 | /* l < c, c > r, l < r - swap c & r to get ordered */ |
917 | qsort_swap(pc_left, u_left); |
918 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
919 | } else if (s == 0) { |
920 | /* l < c, c > r, l == r - swap c&r, grow pc */ |
921 | qsort_swap(pc_left, u_left); |
922 | --pc_left; |
923 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
924 | } else { |
925 | /* l < c, c > r, l > r - make lcr into rlc to get ordered */ |
926 | qsort_rotate(pc_left, u_right, u_left); |
927 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
928 | } |
929 | } |
930 | } else if (s == 0) { |
931 | /* l == c */ |
932 | s = qsort_cmp(pc_left, u_left); |
933 | if (s < 0) { |
934 | /* l == c, c < r - already in order, grow pc */ |
935 | --pc_left; |
936 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
937 | } else if (s == 0) { |
938 | /* l == c, c == r - already in order, grow pc both ways */ |
939 | --pc_left; |
940 | ++pc_right; |
941 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
942 | } else { |
943 | /* l == c, c > r - swap l & r, grow pc */ |
944 | qsort_swap(u_right, u_left); |
945 | ++pc_right; |
946 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
947 | } |
948 | } else { |
949 | /* l > c */ |
950 | s = qsort_cmp(pc_left, u_left); |
951 | if (s < 0) { |
952 | /* l > c, c < r - need to know more */ |
953 | s = qsort_cmp(u_right, u_left); |
954 | if (s < 0) { |
955 | /* l > c, c < r, l < r - swap l & c to get ordered */ |
956 | qsort_swap(u_right, pc_left); |
957 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
958 | } else if (s == 0) { |
959 | /* l > c, c < r, l == r - swap l & c, grow pc */ |
960 | qsort_swap(u_right, pc_left); |
961 | ++pc_right; |
962 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
963 | } else { |
964 | /* l > c, c < r, l > r - rotate lcr into crl to order */ |
965 | qsort_rotate(u_right, pc_left, u_left); |
966 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
967 | } |
968 | } else if (s == 0) { |
969 | /* l > c, c == r - swap ends, grow pc */ |
970 | qsort_swap(u_right, u_left); |
971 | --pc_left; |
972 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
973 | } else { |
974 | /* l > c, c > r - swap ends to get in order */ |
975 | qsort_swap(u_right, u_left); |
976 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
977 | } |
978 | } |
979 | /* We now know the 3 middle elements have been compared and |
980 | arranged in the desired order, so we can shrink the uncompared |
981 | sets on both sides |
982 | */ |
983 | --u_right; |
984 | ++u_left; |
985 | qsort_all_asserts(pc_left, pc_right, u_left, u_right); |
986 | |
987 | /* The above massive nested if was the simple part :-). We now have |
988 | the middle 3 elements ordered and we need to scan through the |
989 | uncompared sets on either side, swapping elements that are on |
990 | the wrong side or simply shuffling equal elements around to get |
991 | all equal elements into the pivot chunk. |
992 | */ |
993 | |
994 | for ( ; ; ) { |
995 | int still_work_on_left; |
996 | int still_work_on_right; |
997 | |
998 | /* Scan the uncompared values on the left. If I find a value |
999 | equal to the pivot value, move it over so it is adjacent to |
1000 | the pivot chunk and expand the pivot chunk. If I find a value |
1001 | less than the pivot value, then just leave it - its already |
1002 | on the correct side of the partition. If I find a greater |
1003 | value, then stop the scan. |
1004 | */ |
1005 | while ((still_work_on_left = (u_right >= part_left))) { |
1006 | s = qsort_cmp(u_right, pc_left); |
1007 | if (s < 0) { |
1008 | --u_right; |
1009 | } else if (s == 0) { |
1010 | --pc_left; |
1011 | if (pc_left != u_right) { |
1012 | qsort_swap(u_right, pc_left); |
1013 | } |
1014 | --u_right; |
1015 | } else { |
1016 | break; |
1017 | } |
1018 | qsort_assert(u_right < pc_left); |
1019 | qsort_assert(pc_left <= pc_right); |
1020 | qsort_assert(qsort_cmp(u_right + 1, pc_left) <= 0); |
1021 | qsort_assert(qsort_cmp(pc_left, pc_right) == 0); |
1022 | } |
1023 | |
1024 | /* Do a mirror image scan of uncompared values on the right |
1025 | */ |
1026 | while ((still_work_on_right = (u_left <= part_right))) { |
1027 | s = qsort_cmp(pc_right, u_left); |
1028 | if (s < 0) { |
1029 | ++u_left; |
1030 | } else if (s == 0) { |
1031 | ++pc_right; |
1032 | if (pc_right != u_left) { |
1033 | qsort_swap(pc_right, u_left); |
1034 | } |
1035 | ++u_left; |
1036 | } else { |
1037 | break; |
1038 | } |
1039 | qsort_assert(u_left > pc_right); |
1040 | qsort_assert(pc_left <= pc_right); |
1041 | qsort_assert(qsort_cmp(pc_right, u_left - 1) <= 0); |
1042 | qsort_assert(qsort_cmp(pc_left, pc_right) == 0); |
1043 | } |
1044 | |
1045 | if (still_work_on_left) { |
1046 | /* I know I have a value on the left side which needs to be |
1047 | on the right side, but I need to know more to decide |
1048 | exactly the best thing to do with it. |
1049 | */ |
1050 | if (still_work_on_right) { |
1051 | /* I know I have values on both side which are out of |
1052 | position. This is a big win because I kill two birds |
1053 | with one swap (so to speak). I can advance the |
1054 | uncompared pointers on both sides after swapping both |
1055 | of them into the right place. |
1056 | */ |
1057 | qsort_swap(u_right, u_left); |
1058 | --u_right; |
1059 | ++u_left; |
1060 | qsort_all_asserts(pc_left, pc_right, u_left, u_right); |
1061 | } else { |
1062 | /* I have an out of position value on the left, but the |
1063 | right is fully scanned, so I "slide" the pivot chunk |
1064 | and any less-than values left one to make room for the |
1065 | greater value over on the right. If the out of position |
1066 | value is immediately adjacent to the pivot chunk (there |
1067 | are no less-than values), I can do that with a swap, |
1068 | otherwise, I have to rotate one of the less than values |
1069 | into the former position of the out of position value |
1070 | and the right end of the pivot chunk into the left end |
1071 | (got all that?). |
1072 | */ |
1073 | --pc_left; |
1074 | if (pc_left == u_right) { |
1075 | qsort_swap(u_right, pc_right); |
1076 | qsort_all_asserts(pc_left, pc_right-1, u_left, u_right-1); |
1077 | } else { |
1078 | qsort_rotate(u_right, pc_left, pc_right); |
1079 | qsort_all_asserts(pc_left, pc_right-1, u_left, u_right-1); |
1080 | } |
1081 | --pc_right; |
1082 | --u_right; |
1083 | } |
1084 | } else if (still_work_on_right) { |
1085 | /* Mirror image of complex case above: I have an out of |
1086 | position value on the right, but the left is fully |
1087 | scanned, so I need to shuffle things around to make room |
1088 | for the right value on the left. |
1089 | */ |
1090 | ++pc_right; |
1091 | if (pc_right == u_left) { |
1092 | qsort_swap(u_left, pc_left); |
1093 | qsort_all_asserts(pc_left+1, pc_right, u_left+1, u_right); |
1094 | } else { |
1095 | qsort_rotate(pc_right, pc_left, u_left); |
1096 | qsort_all_asserts(pc_left+1, pc_right, u_left+1, u_right); |
1097 | } |
1098 | ++pc_left; |
1099 | ++u_left; |
1100 | } else { |
1101 | /* No more scanning required on either side of partition, |
1102 | break out of loop and figure out next set of partitions |
1103 | */ |
1104 | break; |
1105 | } |
1106 | } |
1107 | |
1108 | /* The elements in the pivot chunk are now in the right place. They |
1109 | will never move or be compared again. All I have to do is decide |
1110 | what to do with the stuff to the left and right of the pivot |
1111 | chunk. |
1112 | |
1113 | Notes on the QSORT_ORDER_GUESS ifdef code: |
1114 | |
1115 | 1. If I just built these partitions without swapping any (or |
1116 | very many) elements, there is a chance that the elements are |
1117 | already ordered properly (being properly ordered will |
1118 | certainly result in no swapping, but the converse can't be |
1119 | proved :-). |
1120 | |
1121 | 2. A (properly written) insertion sort will run faster on |
1122 | already ordered data than qsort will. |
1123 | |
1124 | 3. Perhaps there is some way to make a good guess about |
1125 | switching to an insertion sort earlier than partition size 6 |
1126 | (for instance - we could save the partition size on the stack |
1127 | and increase the size each time we find we didn't swap, thus |
1128 | switching to insertion sort earlier for partitions with a |
1129 | history of not swapping). |
1130 | |
1131 | 4. Naturally, if I just switch right away, it will make |
1132 | artificial benchmarks with pure ascending (or descending) |
1133 | data look really good, but is that a good reason in general? |
1134 | Hard to say... |
1135 | */ |
1136 | |
1137 | #ifdef QSORT_ORDER_GUESS |
1138 | if (swapped < 3) { |
1139 | #if QSORT_ORDER_GUESS == 1 |
1140 | qsort_break_even = (part_right - part_left) + 1; |
1141 | #endif |
1142 | #if QSORT_ORDER_GUESS == 2 |
1143 | qsort_break_even *= 2; |
1144 | #endif |
1145 | #if QSORT_ORDER_GUESS == 3 |
1146 | int prev_break = qsort_break_even; |
1147 | qsort_break_even *= qsort_break_even; |
1148 | if (qsort_break_even < prev_break) { |
1149 | qsort_break_even = (part_right - part_left) + 1; |
1150 | } |
1151 | #endif |
1152 | } else { |
1153 | qsort_break_even = QSORT_BREAK_EVEN; |
1154 | } |
1155 | #endif |
1156 | |
1157 | if (part_left < pc_left) { |
1158 | /* There are elements on the left which need more processing. |
1159 | Check the right as well before deciding what to do. |
1160 | */ |
1161 | if (pc_right < part_right) { |
1162 | /* We have two partitions to be sorted. Stack the biggest one |
1163 | and process the smallest one on the next iteration. This |
1164 | minimizes the stack height by insuring that any additional |
1165 | stack entries must come from the smallest partition which |
1166 | (because it is smallest) will have the fewest |
1167 | opportunities to generate additional stack entries. |
1168 | */ |
1169 | if ((part_right - pc_right) > (pc_left - part_left)) { |
1170 | /* stack the right partition, process the left */ |
1171 | partition_stack[next_stack_entry].left = pc_right + 1; |
1172 | partition_stack[next_stack_entry].right = part_right; |
1173 | #ifdef QSORT_ORDER_GUESS |
1174 | partition_stack[next_stack_entry].qsort_break_even = qsort_break_even; |
1175 | #endif |
1176 | part_right = pc_left - 1; |
1177 | } else { |
1178 | /* stack the left partition, process the right */ |
1179 | partition_stack[next_stack_entry].left = part_left; |
1180 | partition_stack[next_stack_entry].right = pc_left - 1; |
1181 | #ifdef QSORT_ORDER_GUESS |
1182 | partition_stack[next_stack_entry].qsort_break_even = qsort_break_even; |
1183 | #endif |
1184 | part_left = pc_right + 1; |
1185 | } |
1186 | qsort_assert(next_stack_entry < QSORT_MAX_STACK); |
1187 | ++next_stack_entry; |
1188 | } else { |
1189 | /* The elements on the left are the only remaining elements |
1190 | that need sorting, arrange for them to be processed as the |
1191 | next partition. |
1192 | */ |
1193 | part_right = pc_left - 1; |
1194 | } |
1195 | } else if (pc_right < part_right) { |
1196 | /* There is only one chunk on the right to be sorted, make it |
1197 | the new partition and loop back around. |
1198 | */ |
1199 | part_left = pc_right + 1; |
1200 | } else { |
1201 | /* This whole partition wound up in the pivot chunk, so |
1202 | we need to get a new partition off the stack. |
1203 | */ |
1204 | if (next_stack_entry == 0) { |
1205 | /* the stack is empty - we are done */ |
1206 | break; |
1207 | } |
1208 | --next_stack_entry; |
1209 | part_left = partition_stack[next_stack_entry].left; |
1210 | part_right = partition_stack[next_stack_entry].right; |
1211 | #ifdef QSORT_ORDER_GUESS |
1212 | qsort_break_even = partition_stack[next_stack_entry].qsort_break_even; |
1213 | #endif |
1214 | } |
1215 | } else { |
1216 | /* This partition is too small to fool with qsort complexity, just |
1217 | do an ordinary insertion sort to minimize overhead. |
1218 | */ |
1219 | int i; |
1220 | /* Assume 1st element is in right place already, and start checking |
1221 | at 2nd element to see where it should be inserted. |
1222 | */ |
1223 | for (i = part_left + 1; i <= part_right; ++i) { |
1224 | int j; |
1225 | /* Scan (backwards - just in case 'i' is already in right place) |
1226 | through the elements already sorted to see if the ith element |
1227 | belongs ahead of one of them. |
1228 | */ |
1229 | for (j = i - 1; j >= part_left; --j) { |
1230 | if (qsort_cmp(i, j) >= 0) { |
1231 | /* i belongs right after j |
1232 | */ |
1233 | break; |
1234 | } |
1235 | } |
1236 | ++j; |
1237 | if (j != i) { |
1238 | /* Looks like we really need to move some things |
1239 | */ |
1240 | int k; |
1241 | temp = array[i]; |
1242 | for (k = i - 1; k >= j; --k) |
1243 | array[k + 1] = array[k]; |
1244 | array[j] = temp; |
1245 | } |
1246 | } |
1247 | |
1248 | /* That partition is now sorted, grab the next one, or get out |
1249 | of the loop if there aren't any more. |
1250 | */ |
1251 | |
1252 | if (next_stack_entry == 0) { |
1253 | /* the stack is empty - we are done */ |
1254 | break; |
1255 | } |
1256 | --next_stack_entry; |
1257 | part_left = partition_stack[next_stack_entry].left; |
1258 | part_right = partition_stack[next_stack_entry].right; |
1259 | #ifdef QSORT_ORDER_GUESS |
1260 | qsort_break_even = partition_stack[next_stack_entry].qsort_break_even; |
1261 | #endif |
1262 | } |
1263 | } |
1264 | |
1265 | /* Believe it or not, the array is sorted at this point! */ |
1266 | } |
1267 | |
84d4ea48 |
1268 | /* Stabilize what is, presumably, an otherwise unstable sort method. |
1269 | * We do that by allocating (or having on hand) an array of pointers |
1270 | * that is the same size as the original array of elements to be sorted. |
1271 | * We initialize this parallel array with the addresses of the original |
1272 | * array elements. This indirection can make you crazy. |
1273 | * Some pictures can help. After initializing, we have |
1274 | * |
1275 | * indir list1 |
1276 | * +----+ +----+ |
1277 | * | | --------------> | | ------> first element to be sorted |
1278 | * +----+ +----+ |
1279 | * | | --------------> | | ------> second element to be sorted |
1280 | * +----+ +----+ |
1281 | * | | --------------> | | ------> third element to be sorted |
1282 | * +----+ +----+ |
1283 | * ... |
1284 | * +----+ +----+ |
1285 | * | | --------------> | | ------> n-1st element to be sorted |
1286 | * +----+ +----+ |
1287 | * | | --------------> | | ------> n-th element to be sorted |
1288 | * +----+ +----+ |
1289 | * |
1290 | * During the sort phase, we leave the elements of list1 where they are, |
1291 | * and sort the pointers in the indirect array in the same order determined |
1292 | * by the original comparison routine on the elements pointed to. |
1293 | * Because we don't move the elements of list1 around through |
1294 | * this phase, we can break ties on elements that compare equal |
1295 | * using their address in the list1 array, ensuring stabilty. |
1296 | * This leaves us with something looking like |
1297 | * |
1298 | * indir list1 |
1299 | * +----+ +----+ |
1300 | * | | --+ +---> | | ------> first element to be sorted |
1301 | * +----+ | | +----+ |
1302 | * | | --|-------|---> | | ------> second element to be sorted |
1303 | * +----+ | | +----+ |
1304 | * | | --|-------+ +-> | | ------> third element to be sorted |
1305 | * +----+ | | +----+ |
1306 | * ... |
1307 | * +----+ | | | | +----+ |
1308 | * | | ---|-+ | +--> | | ------> n-1st element to be sorted |
1309 | * +----+ | | +----+ |
1310 | * | | ---+ +----> | | ------> n-th element to be sorted |
1311 | * +----+ +----+ |
1312 | * |
1313 | * where the i-th element of the indirect array points to the element |
1314 | * that should be i-th in the sorted array. After the sort phase, |
1315 | * we have to put the elements of list1 into the places |
1316 | * dictated by the indirect array. |
1317 | */ |
1318 | |
84d4ea48 |
1319 | |
1320 | static I32 |
1321 | cmpindir(pTHX_ gptr a, gptr b) |
1322 | { |
1323 | I32 sense; |
1324 | gptr *ap = (gptr *)a; |
1325 | gptr *bp = (gptr *)b; |
1326 | |
147f47de |
1327 | if ((sense = PL_sort_RealCmp(aTHX_ *ap, *bp)) == 0) |
84d4ea48 |
1328 | sense = (ap > bp) ? 1 : ((ap < bp) ? -1 : 0); |
1329 | return sense; |
1330 | } |
1331 | |
6c3fb703 |
1332 | static I32 |
1333 | cmpindir_desc(pTHX_ gptr a, gptr b) |
1334 | { |
1335 | I32 sense; |
1336 | gptr *ap = (gptr *)a; |
1337 | gptr *bp = (gptr *)b; |
1338 | |
1339 | /* Reverse the default */ |
1340 | if ((sense = PL_sort_RealCmp(aTHX_ *ap, *bp))) |
1341 | return -sense; |
1342 | /* But don't reverse the stability test. */ |
1343 | return (ap > bp) ? 1 : ((ap < bp) ? -1 : 0); |
1344 | |
1345 | } |
1346 | |
84d4ea48 |
1347 | STATIC void |
6c3fb703 |
1348 | S_qsortsv(pTHX_ gptr *list1, size_t nmemb, SVCOMPARE_t cmp, U32 flags) |
84d4ea48 |
1349 | { |
045ac317 |
1350 | SV *hintsv; |
84d4ea48 |
1351 | |
045ac317 |
1352 | if (SORTHINTS(hintsv) & HINT_SORT_STABLE) { |
84d4ea48 |
1353 | register gptr **pp, *q; |
1354 | register size_t n, j, i; |
1355 | gptr *small[SMALLSORT], **indir, tmp; |
1356 | SVCOMPARE_t savecmp; |
1357 | if (nmemb <= 1) return; /* sorted trivially */ |
4eb872f6 |
1358 | |
84d4ea48 |
1359 | /* Small arrays can use the stack, big ones must be allocated */ |
1360 | if (nmemb <= SMALLSORT) indir = small; |
1361 | else { New(1799, indir, nmemb, gptr *); } |
4eb872f6 |
1362 | |
84d4ea48 |
1363 | /* Copy pointers to original array elements into indirect array */ |
1364 | for (n = nmemb, pp = indir, q = list1; n--; ) *pp++ = q++; |
4eb872f6 |
1365 | |
147f47de |
1366 | savecmp = PL_sort_RealCmp; /* Save current comparison routine, if any */ |
1367 | PL_sort_RealCmp = cmp; /* Put comparison routine where cmpindir can find it */ |
4eb872f6 |
1368 | |
84d4ea48 |
1369 | /* sort, with indirection */ |
6c3fb703 |
1370 | S_qsortsvu(aTHX_ (gptr *)indir, nmemb, |
1371 | flags ? cmpindir_desc : cmpindir); |
4eb872f6 |
1372 | |
84d4ea48 |
1373 | pp = indir; |
1374 | q = list1; |
1375 | for (n = nmemb; n--; ) { |
1376 | /* Assert A: all elements of q with index > n are already |
1377 | * in place. This is vacuosly true at the start, and we |
1378 | * put element n where it belongs below (if it wasn't |
1379 | * already where it belonged). Assert B: we only move |
1380 | * elements that aren't where they belong, |
1381 | * so, by A, we never tamper with elements above n. |
1382 | */ |
1383 | j = pp[n] - q; /* This sets j so that q[j] is |
1384 | * at pp[n]. *pp[j] belongs in |
1385 | * q[j], by construction. |
1386 | */ |
1387 | if (n != j) { /* all's well if n == j */ |
1388 | tmp = q[j]; /* save what's in q[j] */ |
1389 | do { |
1390 | q[j] = *pp[j]; /* put *pp[j] where it belongs */ |
1391 | i = pp[j] - q; /* the index in q of the element |
1392 | * just moved */ |
1393 | pp[j] = q + j; /* this is ok now */ |
1394 | } while ((j = i) != n); |
1395 | /* There are only finitely many (nmemb) addresses |
1396 | * in the pp array. |
1397 | * So we must eventually revisit an index we saw before. |
1398 | * Suppose the first revisited index is k != n. |
1399 | * An index is visited because something else belongs there. |
1400 | * If we visit k twice, then two different elements must |
1401 | * belong in the same place, which cannot be. |
1402 | * So j must get back to n, the loop terminates, |
1403 | * and we put the saved element where it belongs. |
1404 | */ |
1405 | q[n] = tmp; /* put what belongs into |
1406 | * the n-th element */ |
1407 | } |
1408 | } |
1409 | |
1410 | /* free iff allocated */ |
1411 | if (indir != small) { Safefree(indir); } |
1412 | /* restore prevailing comparison routine */ |
147f47de |
1413 | PL_sort_RealCmp = savecmp; |
6c3fb703 |
1414 | } else if (flags) { |
1415 | SVCOMPARE_t savecmp = PL_sort_RealCmp; /* Save current comparison routine, if any */ |
1416 | PL_sort_RealCmp = cmp; /* Put comparison routine where cmp_desc can find it */ |
1417 | cmp = cmp_desc; |
1418 | S_qsortsvu(aTHX_ list1, nmemb, cmp); |
1419 | /* restore prevailing comparison routine */ |
1420 | PL_sort_RealCmp = savecmp; |
c53fc8a6 |
1421 | } else { |
1422 | S_qsortsvu(aTHX_ list1, nmemb, cmp); |
84d4ea48 |
1423 | } |
1424 | } |
4eb872f6 |
1425 | |
1426 | /* |
ccfc67b7 |
1427 | =head1 Array Manipulation Functions |
1428 | |
84d4ea48 |
1429 | =for apidoc sortsv |
1430 | |
1431 | Sort an array. Here is an example: |
1432 | |
4eb872f6 |
1433 | sortsv(AvARRAY(av), av_len(av)+1, Perl_sv_cmp_locale); |
84d4ea48 |
1434 | |
78210658 |
1435 | See lib/sort.pm for details about controlling the sorting algorithm. |
1436 | |
84d4ea48 |
1437 | =cut |
1438 | */ |
4eb872f6 |
1439 | |
84d4ea48 |
1440 | void |
1441 | Perl_sortsv(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp) |
1442 | { |
6c3fb703 |
1443 | void (*sortsvp)(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp, U32 flags) |
1444 | = S_mergesortsv; |
045ac317 |
1445 | SV *hintsv; |
84d4ea48 |
1446 | I32 hints; |
4eb872f6 |
1447 | |
78210658 |
1448 | /* Sun's Compiler (cc: WorkShop Compilers 4.2 30 Oct 1996 C 4.2) used |
1449 | to miscompile this function under optimization -O. If you get test |
1450 | errors related to picking the correct sort() function, try recompiling |
1451 | this file without optimiziation. -- A.D. 4/2002. |
1452 | */ |
045ac317 |
1453 | hints = SORTHINTS(hintsv); |
78210658 |
1454 | if (hints & HINT_SORT_QUICKSORT) { |
1455 | sortsvp = S_qsortsv; |
1456 | } |
1457 | else { |
1458 | /* The default as of 5.8.0 is mergesort */ |
1459 | sortsvp = S_mergesortsv; |
84d4ea48 |
1460 | } |
4eb872f6 |
1461 | |
6c3fb703 |
1462 | sortsvp(aTHX_ array, nmemb, cmp, 0); |
1463 | } |
1464 | |
1465 | |
1466 | void |
1467 | S_sortsv_desc(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp) |
1468 | { |
1469 | void (*sortsvp)(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp, U32 flags) |
1470 | = S_mergesortsv; |
1471 | SV *hintsv; |
1472 | I32 hints; |
1473 | |
1474 | /* Sun's Compiler (cc: WorkShop Compilers 4.2 30 Oct 1996 C 4.2) used |
1475 | to miscompile this function under optimization -O. If you get test |
1476 | errors related to picking the correct sort() function, try recompiling |
1477 | this file without optimiziation. -- A.D. 4/2002. |
1478 | */ |
1479 | hints = SORTHINTS(hintsv); |
1480 | if (hints & HINT_SORT_QUICKSORT) { |
1481 | sortsvp = S_qsortsv; |
1482 | } |
1483 | else { |
1484 | /* The default as of 5.8.0 is mergesort */ |
1485 | sortsvp = S_mergesortsv; |
1486 | } |
1487 | |
1488 | sortsvp(aTHX_ array, nmemb, cmp, 1); |
84d4ea48 |
1489 | } |
1490 | |
1491 | PP(pp_sort) |
1492 | { |
1493 | dSP; dMARK; dORIGMARK; |
fe1bc4cf |
1494 | register SV **p1 = ORIGMARK+1, **p2; |
1495 | register I32 max, i; |
1496 | AV* av = Nullav; |
84d4ea48 |
1497 | HV *stash; |
1498 | GV *gv; |
1499 | CV *cv = 0; |
1500 | I32 gimme = GIMME; |
1501 | OP* nextop = PL_op->op_next; |
1502 | I32 overloading = 0; |
1503 | bool hasargs = FALSE; |
1504 | I32 is_xsub = 0; |
fe1bc4cf |
1505 | I32 sorting_av = 0; |
471178c0 |
1506 | U8 private = PL_op->op_private; |
1507 | U8 flags = PL_op->op_flags; |
6c3fb703 |
1508 | void (*sortsvp)(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp) |
1509 | = Perl_sortsv; |
84d4ea48 |
1510 | |
1511 | if (gimme != G_ARRAY) { |
1512 | SP = MARK; |
1513 | RETPUSHUNDEF; |
1514 | } |
1515 | |
1516 | ENTER; |
1517 | SAVEVPTR(PL_sortcop); |
471178c0 |
1518 | if (flags & OPf_STACKED) { |
1519 | if (flags & OPf_SPECIAL) { |
84d4ea48 |
1520 | OP *kid = cLISTOP->op_first->op_sibling; /* pass pushmark */ |
1521 | kid = kUNOP->op_first; /* pass rv2gv */ |
1522 | kid = kUNOP->op_first; /* pass leave */ |
1523 | PL_sortcop = kid->op_next; |
1524 | stash = CopSTASH(PL_curcop); |
1525 | } |
1526 | else { |
1527 | cv = sv_2cv(*++MARK, &stash, &gv, 0); |
1528 | if (cv && SvPOK(cv)) { |
1529 | STRLEN n_a; |
1530 | char *proto = SvPV((SV*)cv, n_a); |
1531 | if (proto && strEQ(proto, "$$")) { |
1532 | hasargs = TRUE; |
1533 | } |
1534 | } |
1535 | if (!(cv && CvROOT(cv))) { |
1536 | if (cv && CvXSUB(cv)) { |
1537 | is_xsub = 1; |
1538 | } |
1539 | else if (gv) { |
1540 | SV *tmpstr = sv_newmortal(); |
1541 | gv_efullname3(tmpstr, gv, Nullch); |
35c1215d |
1542 | DIE(aTHX_ "Undefined sort subroutine \"%"SVf"\" called", |
1543 | tmpstr); |
84d4ea48 |
1544 | } |
1545 | else { |
1546 | DIE(aTHX_ "Undefined subroutine in sort"); |
1547 | } |
1548 | } |
1549 | |
1550 | if (is_xsub) |
1551 | PL_sortcop = (OP*)cv; |
1552 | else { |
1553 | PL_sortcop = CvSTART(cv); |
1554 | SAVEVPTR(CvROOT(cv)->op_ppaddr); |
1555 | CvROOT(cv)->op_ppaddr = PL_ppaddr[OP_NULL]; |
1556 | |
dd2155a4 |
1557 | PAD_SET_CUR(CvPADLIST(cv), 1); |
84d4ea48 |
1558 | } |
1559 | } |
1560 | } |
1561 | else { |
1562 | PL_sortcop = Nullop; |
1563 | stash = CopSTASH(PL_curcop); |
1564 | } |
1565 | |
fe1bc4cf |
1566 | /* optimiser converts "@a = sort @a" to "sort \@a"; |
1567 | * in case of tied @a, pessimise: push (@a) onto stack, then assign |
1568 | * result back to @a at the end of this function */ |
471178c0 |
1569 | if (private & OPpSORT_INPLACE) { |
fe1bc4cf |
1570 | assert( MARK+1 == SP && *SP && SvTYPE(*SP) == SVt_PVAV); |
1571 | (void)POPMARK; /* remove mark associated with ex-OP_AASSIGN */ |
1572 | av = (AV*)(*SP); |
1573 | max = AvFILL(av) + 1; |
1574 | if (SvMAGICAL(av)) { |
1575 | MEXTEND(SP, max); |
1576 | p2 = SP; |
1577 | for (i=0; i < (U32)max; i++) { |
1578 | SV **svp = av_fetch(av, i, FALSE); |
1579 | *SP++ = (svp) ? *svp : Nullsv; |
1580 | } |
1581 | } |
1582 | else { |
1583 | p1 = p2 = AvARRAY(av); |
1584 | sorting_av = 1; |
1585 | } |
1586 | } |
1587 | else { |
1588 | p2 = MARK+1; |
1589 | max = SP - MARK; |
1590 | } |
1591 | |
471178c0 |
1592 | if (private & OPpSORT_DESCEND) { |
6c3fb703 |
1593 | sortsvp = S_sortsv_desc; |
1594 | } |
1595 | |
fe1bc4cf |
1596 | /* shuffle stack down, removing optional initial cv (p1!=p2), plus any |
1597 | * nulls; also stringify any args */ |
1598 | for (i=max; i > 0 ; i--) { |
1599 | if ((*p1 = *p2++)) { /* Weed out nulls. */ |
1600 | SvTEMP_off(*p1); |
1601 | if (!PL_sortcop && !SvPOK(*p1)) { |
84d4ea48 |
1602 | STRLEN n_a; |
fe1bc4cf |
1603 | if (SvAMAGIC(*p1)) |
84d4ea48 |
1604 | overloading = 1; |
1605 | else |
fe1bc4cf |
1606 | (void)sv_2pv(*p1, &n_a); |
84d4ea48 |
1607 | } |
fe1bc4cf |
1608 | p1++; |
84d4ea48 |
1609 | } |
fe1bc4cf |
1610 | else |
1611 | max--; |
84d4ea48 |
1612 | } |
fe1bc4cf |
1613 | if (sorting_av) |
1614 | AvFILLp(av) = max-1; |
1615 | |
1616 | if (max > 1) { |
471178c0 |
1617 | SV **start; |
fe1bc4cf |
1618 | if (PL_sortcop) { |
84d4ea48 |
1619 | PERL_CONTEXT *cx; |
1620 | SV** newsp; |
1621 | bool oldcatch = CATCH_GET; |
1622 | |
1623 | SAVETMPS; |
1624 | SAVEOP(); |
1625 | |
1626 | CATCH_SET(TRUE); |
1627 | PUSHSTACKi(PERLSI_SORT); |
1628 | if (!hasargs && !is_xsub) { |
1629 | if (PL_sortstash != stash || !PL_firstgv || !PL_secondgv) { |
1630 | SAVESPTR(PL_firstgv); |
1631 | SAVESPTR(PL_secondgv); |
1632 | PL_firstgv = gv_fetchpv("a", TRUE, SVt_PV); |
1633 | PL_secondgv = gv_fetchpv("b", TRUE, SVt_PV); |
1634 | PL_sortstash = stash; |
1635 | } |
84d4ea48 |
1636 | SAVESPTR(GvSV(PL_firstgv)); |
1637 | SAVESPTR(GvSV(PL_secondgv)); |
1638 | } |
1639 | |
1640 | PUSHBLOCK(cx, CXt_NULL, PL_stack_base); |
471178c0 |
1641 | if (!(flags & OPf_SPECIAL)) { |
84d4ea48 |
1642 | cx->cx_type = CXt_SUB; |
1643 | cx->blk_gimme = G_SCALAR; |
1644 | PUSHSUB(cx); |
84d4ea48 |
1645 | } |
1646 | PL_sortcxix = cxstack_ix; |
1647 | |
1648 | if (hasargs && !is_xsub) { |
1649 | /* This is mostly copied from pp_entersub */ |
dd2155a4 |
1650 | AV *av = (AV*)PAD_SVl(0); |
84d4ea48 |
1651 | |
84d4ea48 |
1652 | cx->blk_sub.savearray = GvAV(PL_defgv); |
1653 | GvAV(PL_defgv) = (AV*)SvREFCNT_inc(av); |
dd2155a4 |
1654 | CX_CURPAD_SAVE(cx->blk_sub); |
84d4ea48 |
1655 | cx->blk_sub.argarray = av; |
1656 | } |
471178c0 |
1657 | |
1658 | start = p1 - max; |
1659 | sortsvp(aTHX_ start, max, |
1660 | is_xsub ? sortcv_xsub : hasargs ? sortcv_stacked : sortcv); |
84d4ea48 |
1661 | |
1662 | POPBLOCK(cx,PL_curpm); |
1663 | PL_stack_sp = newsp; |
1664 | POPSTACK; |
1665 | CATCH_SET(oldcatch); |
1666 | } |
fe1bc4cf |
1667 | else { |
84d4ea48 |
1668 | MEXTEND(SP, 20); /* Can't afford stack realloc on signal. */ |
471178c0 |
1669 | start = sorting_av ? AvARRAY(av) : ORIGMARK+1; |
1670 | sortsvp(aTHX_ start, max, |
1671 | (private & OPpSORT_NUMERIC) |
1672 | ? ( (private & OPpSORT_INTEGER) |
84d4ea48 |
1673 | ? ( overloading ? amagic_i_ncmp : sv_i_ncmp) |
1674 | : ( overloading ? amagic_ncmp : sv_ncmp)) |
1675 | : ( IN_LOCALE_RUNTIME |
1676 | ? ( overloading |
1677 | ? amagic_cmp_locale |
1678 | : sv_cmp_locale_static) |
1679 | : ( overloading ? amagic_cmp : sv_cmp_static))); |
471178c0 |
1680 | } |
1681 | if (private & OPpSORT_REVERSE) { |
1682 | SV **q = start+max-1; |
1683 | while (start < q) { |
1684 | SV *tmp = *start; |
1685 | *start++ = *q; |
1686 | *q-- = tmp; |
84d4ea48 |
1687 | } |
1688 | } |
1689 | } |
fe1bc4cf |
1690 | if (av && !sorting_av) { |
1691 | /* simulate pp_aassign of tied AV */ |
1692 | SV *sv; |
1693 | SV** base, **didstore; |
1694 | for (base = ORIGMARK+1, i=0; i < max; i++) { |
f2b990bf |
1695 | sv = newSVsv(base[i]); |
fe1bc4cf |
1696 | base[i] = sv; |
1697 | } |
1698 | av_clear(av); |
1699 | av_extend(av, max); |
1700 | for (i=0; i < max; i++) { |
1701 | sv = base[i]; |
1702 | didstore = av_store(av, i, sv); |
1703 | if (SvSMAGICAL(sv)) |
1704 | mg_set(sv); |
1705 | if (!didstore) |
1706 | sv_2mortal(sv); |
1707 | } |
1708 | } |
84d4ea48 |
1709 | LEAVE; |
fe1bc4cf |
1710 | PL_stack_sp = ORIGMARK + (sorting_av ? 0 : max); |
84d4ea48 |
1711 | return nextop; |
1712 | } |
1713 | |
1714 | static I32 |
1715 | sortcv(pTHX_ SV *a, SV *b) |
1716 | { |
1717 | I32 oldsaveix = PL_savestack_ix; |
1718 | I32 oldscopeix = PL_scopestack_ix; |
1719 | I32 result; |
1720 | GvSV(PL_firstgv) = a; |
1721 | GvSV(PL_secondgv) = b; |
1722 | PL_stack_sp = PL_stack_base; |
1723 | PL_op = PL_sortcop; |
1724 | CALLRUNOPS(aTHX); |
1725 | if (PL_stack_sp != PL_stack_base + 1) |
1726 | Perl_croak(aTHX_ "Sort subroutine didn't return single value"); |
1727 | if (!SvNIOKp(*PL_stack_sp)) |
1728 | Perl_croak(aTHX_ "Sort subroutine didn't return a numeric value"); |
1729 | result = SvIV(*PL_stack_sp); |
1730 | while (PL_scopestack_ix > oldscopeix) { |
1731 | LEAVE; |
1732 | } |
1733 | leave_scope(oldsaveix); |
1734 | return result; |
1735 | } |
1736 | |
1737 | static I32 |
1738 | sortcv_stacked(pTHX_ SV *a, SV *b) |
1739 | { |
1740 | I32 oldsaveix = PL_savestack_ix; |
1741 | I32 oldscopeix = PL_scopestack_ix; |
1742 | I32 result; |
1743 | AV *av; |
1744 | |
84d4ea48 |
1745 | av = GvAV(PL_defgv); |
84d4ea48 |
1746 | |
1747 | if (AvMAX(av) < 1) { |
1748 | SV** ary = AvALLOC(av); |
1749 | if (AvARRAY(av) != ary) { |
1750 | AvMAX(av) += AvARRAY(av) - AvALLOC(av); |
1751 | SvPVX(av) = (char*)ary; |
1752 | } |
1753 | if (AvMAX(av) < 1) { |
1754 | AvMAX(av) = 1; |
1755 | Renew(ary,2,SV*); |
1756 | SvPVX(av) = (char*)ary; |
1757 | } |
1758 | } |
1759 | AvFILLp(av) = 1; |
1760 | |
1761 | AvARRAY(av)[0] = a; |
1762 | AvARRAY(av)[1] = b; |
1763 | PL_stack_sp = PL_stack_base; |
1764 | PL_op = PL_sortcop; |
1765 | CALLRUNOPS(aTHX); |
1766 | if (PL_stack_sp != PL_stack_base + 1) |
1767 | Perl_croak(aTHX_ "Sort subroutine didn't return single value"); |
1768 | if (!SvNIOKp(*PL_stack_sp)) |
1769 | Perl_croak(aTHX_ "Sort subroutine didn't return a numeric value"); |
1770 | result = SvIV(*PL_stack_sp); |
1771 | while (PL_scopestack_ix > oldscopeix) { |
1772 | LEAVE; |
1773 | } |
1774 | leave_scope(oldsaveix); |
1775 | return result; |
1776 | } |
1777 | |
1778 | static I32 |
1779 | sortcv_xsub(pTHX_ SV *a, SV *b) |
1780 | { |
1781 | dSP; |
1782 | I32 oldsaveix = PL_savestack_ix; |
1783 | I32 oldscopeix = PL_scopestack_ix; |
1784 | I32 result; |
1785 | CV *cv=(CV*)PL_sortcop; |
1786 | |
1787 | SP = PL_stack_base; |
1788 | PUSHMARK(SP); |
1789 | EXTEND(SP, 2); |
1790 | *++SP = a; |
1791 | *++SP = b; |
1792 | PUTBACK; |
1793 | (void)(*CvXSUB(cv))(aTHX_ cv); |
1794 | if (PL_stack_sp != PL_stack_base + 1) |
1795 | Perl_croak(aTHX_ "Sort subroutine didn't return single value"); |
1796 | if (!SvNIOKp(*PL_stack_sp)) |
1797 | Perl_croak(aTHX_ "Sort subroutine didn't return a numeric value"); |
1798 | result = SvIV(*PL_stack_sp); |
1799 | while (PL_scopestack_ix > oldscopeix) { |
1800 | LEAVE; |
1801 | } |
1802 | leave_scope(oldsaveix); |
1803 | return result; |
1804 | } |
1805 | |
1806 | |
1807 | static I32 |
1808 | sv_ncmp(pTHX_ SV *a, SV *b) |
1809 | { |
1810 | NV nv1 = SvNV(a); |
1811 | NV nv2 = SvNV(b); |
1812 | return nv1 < nv2 ? -1 : nv1 > nv2 ? 1 : 0; |
1813 | } |
1814 | |
1815 | static I32 |
1816 | sv_i_ncmp(pTHX_ SV *a, SV *b) |
1817 | { |
1818 | IV iv1 = SvIV(a); |
1819 | IV iv2 = SvIV(b); |
1820 | return iv1 < iv2 ? -1 : iv1 > iv2 ? 1 : 0; |
1821 | } |
1822 | #define tryCALL_AMAGICbin(left,right,meth,svp) STMT_START { \ |
1823 | *svp = Nullsv; \ |
1824 | if (PL_amagic_generation) { \ |
1825 | if (SvAMAGIC(left)||SvAMAGIC(right))\ |
1826 | *svp = amagic_call(left, \ |
1827 | right, \ |
1828 | CAT2(meth,_amg), \ |
1829 | 0); \ |
1830 | } \ |
1831 | } STMT_END |
1832 | |
1833 | static I32 |
1834 | amagic_ncmp(pTHX_ register SV *a, register SV *b) |
1835 | { |
1836 | SV *tmpsv; |
1837 | tryCALL_AMAGICbin(a,b,ncmp,&tmpsv); |
1838 | if (tmpsv) { |
1839 | NV d; |
4eb872f6 |
1840 | |
84d4ea48 |
1841 | if (SvIOK(tmpsv)) { |
1842 | I32 i = SvIVX(tmpsv); |
1843 | if (i > 0) |
1844 | return 1; |
1845 | return i? -1 : 0; |
1846 | } |
1847 | d = SvNV(tmpsv); |
1848 | if (d > 0) |
1849 | return 1; |
1850 | return d? -1 : 0; |
1851 | } |
1852 | return sv_ncmp(aTHX_ a, b); |
1853 | } |
1854 | |
1855 | static I32 |
1856 | amagic_i_ncmp(pTHX_ register SV *a, register SV *b) |
1857 | { |
1858 | SV *tmpsv; |
1859 | tryCALL_AMAGICbin(a,b,ncmp,&tmpsv); |
1860 | if (tmpsv) { |
1861 | NV d; |
4eb872f6 |
1862 | |
84d4ea48 |
1863 | if (SvIOK(tmpsv)) { |
1864 | I32 i = SvIVX(tmpsv); |
1865 | if (i > 0) |
1866 | return 1; |
1867 | return i? -1 : 0; |
1868 | } |
1869 | d = SvNV(tmpsv); |
1870 | if (d > 0) |
1871 | return 1; |
1872 | return d? -1 : 0; |
1873 | } |
1874 | return sv_i_ncmp(aTHX_ a, b); |
1875 | } |
1876 | |
1877 | static I32 |
1878 | amagic_cmp(pTHX_ register SV *str1, register SV *str2) |
1879 | { |
1880 | SV *tmpsv; |
1881 | tryCALL_AMAGICbin(str1,str2,scmp,&tmpsv); |
1882 | if (tmpsv) { |
1883 | NV d; |
4eb872f6 |
1884 | |
84d4ea48 |
1885 | if (SvIOK(tmpsv)) { |
1886 | I32 i = SvIVX(tmpsv); |
1887 | if (i > 0) |
1888 | return 1; |
1889 | return i? -1 : 0; |
1890 | } |
1891 | d = SvNV(tmpsv); |
1892 | if (d > 0) |
1893 | return 1; |
1894 | return d? -1 : 0; |
1895 | } |
1896 | return sv_cmp(str1, str2); |
1897 | } |
1898 | |
1899 | static I32 |
1900 | amagic_cmp_locale(pTHX_ register SV *str1, register SV *str2) |
1901 | { |
1902 | SV *tmpsv; |
1903 | tryCALL_AMAGICbin(str1,str2,scmp,&tmpsv); |
1904 | if (tmpsv) { |
1905 | NV d; |
4eb872f6 |
1906 | |
84d4ea48 |
1907 | if (SvIOK(tmpsv)) { |
1908 | I32 i = SvIVX(tmpsv); |
1909 | if (i > 0) |
1910 | return 1; |
1911 | return i? -1 : 0; |
1912 | } |
1913 | d = SvNV(tmpsv); |
1914 | if (d > 0) |
1915 | return 1; |
1916 | return d? -1 : 0; |
1917 | } |
1918 | return sv_cmp_locale(str1, str2); |
1919 | } |
241d1a3b |
1920 | |
1921 | /* |
1922 | * Local variables: |
1923 | * c-indentation-style: bsd |
1924 | * c-basic-offset: 4 |
1925 | * indent-tabs-mode: t |
1926 | * End: |
1927 | * |
edf815fd |
1928 | * vim: shiftwidth=4: |
241d1a3b |
1929 | */ |