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 | |
5fe61d93 |
49 | #define dSORTHINTS SV *hintsv = GvSV(gv_fetchpv("sort::hints", GV_ADDMULTI, SVt_IV)) |
50 | #define SORTHINTS (SvIOK(hintsv) ? ((I32)SvIV(hintsv)) : 0) |
84d4ea48 |
51 | |
c53fc8a6 |
52 | #ifndef SMALLSORT |
53 | #define SMALLSORT (200) |
54 | #endif |
55 | |
84d4ea48 |
56 | /* |
57 | * The mergesort implementation is by Peter M. Mcilroy <pmcilroy@lucent.com>. |
58 | * |
59 | * The original code was written in conjunction with BSD Computer Software |
60 | * Research Group at University of California, Berkeley. |
61 | * |
62 | * See also: "Optimistic Merge Sort" (SODA '92) |
63 | * |
64 | * The integration to Perl is by John P. Linderman <jpl@research.att.com>. |
65 | * |
66 | * The code can be distributed under the same terms as Perl itself. |
67 | * |
68 | */ |
69 | |
84d4ea48 |
70 | |
71 | typedef char * aptr; /* pointer for arithmetic on sizes */ |
72 | typedef SV * gptr; /* pointers in our lists */ |
73 | |
74 | /* Binary merge internal sort, with a few special mods |
75 | ** for the special perl environment it now finds itself in. |
76 | ** |
77 | ** Things that were once options have been hotwired |
78 | ** to values suitable for this use. In particular, we'll always |
79 | ** initialize looking for natural runs, we'll always produce stable |
80 | ** output, and we'll always do Peter McIlroy's binary merge. |
81 | */ |
82 | |
83 | /* Pointer types for arithmetic and storage and convenience casts */ |
84 | |
85 | #define APTR(P) ((aptr)(P)) |
86 | #define GPTP(P) ((gptr *)(P)) |
87 | #define GPPP(P) ((gptr **)(P)) |
88 | |
89 | |
90 | /* byte offset from pointer P to (larger) pointer Q */ |
91 | #define BYTEOFF(P, Q) (APTR(Q) - APTR(P)) |
92 | |
93 | #define PSIZE sizeof(gptr) |
94 | |
95 | /* If PSIZE is power of 2, make PSHIFT that power, if that helps */ |
96 | |
97 | #ifdef PSHIFT |
98 | #define PNELEM(P, Q) (BYTEOFF(P,Q) >> (PSHIFT)) |
99 | #define PNBYTE(N) ((N) << (PSHIFT)) |
100 | #define PINDEX(P, N) (GPTP(APTR(P) + PNBYTE(N))) |
101 | #else |
102 | /* Leave optimization to compiler */ |
103 | #define PNELEM(P, Q) (GPTP(Q) - GPTP(P)) |
104 | #define PNBYTE(N) ((N) * (PSIZE)) |
105 | #define PINDEX(P, N) (GPTP(P) + (N)) |
106 | #endif |
107 | |
108 | /* Pointer into other corresponding to pointer into this */ |
109 | #define POTHER(P, THIS, OTHER) GPTP(APTR(OTHER) + BYTEOFF(THIS,P)) |
110 | |
111 | #define FROMTOUPTO(src, dst, lim) do *dst++ = *src++; while(src<lim) |
112 | |
113 | |
114 | /* Runs are identified by a pointer in the auxilliary list. |
115 | ** The pointer is at the start of the list, |
116 | ** and it points to the start of the next list. |
117 | ** NEXT is used as an lvalue, too. |
118 | */ |
119 | |
120 | #define NEXT(P) (*GPPP(P)) |
121 | |
122 | |
123 | /* PTHRESH is the minimum number of pairs with the same sense to justify |
124 | ** checking for a run and extending it. Note that PTHRESH counts PAIRS, |
125 | ** not just elements, so PTHRESH == 8 means a run of 16. |
126 | */ |
127 | |
128 | #define PTHRESH (8) |
129 | |
130 | /* RTHRESH is the number of elements in a run that must compare low |
131 | ** to the low element from the opposing run before we justify |
132 | ** doing a binary rampup instead of single stepping. |
133 | ** In random input, N in a row low should only happen with |
134 | ** probability 2^(1-N), so we can risk that we are dealing |
135 | ** with orderly input without paying much when we aren't. |
136 | */ |
137 | |
138 | #define RTHRESH (6) |
139 | |
140 | |
141 | /* |
142 | ** Overview of algorithm and variables. |
143 | ** The array of elements at list1 will be organized into runs of length 2, |
144 | ** or runs of length >= 2 * PTHRESH. We only try to form long runs when |
145 | ** PTHRESH adjacent pairs compare in the same way, suggesting overall order. |
146 | ** |
147 | ** Unless otherwise specified, pair pointers address the first of two elements. |
148 | ** |
a0288114 |
149 | ** b and b+1 are a pair that compare with sense "sense". |
150 | ** b is the "bottom" of adjacent pairs that might form a longer run. |
84d4ea48 |
151 | ** |
152 | ** p2 parallels b in the list2 array, where runs are defined by |
153 | ** a pointer chain. |
154 | ** |
a0288114 |
155 | ** t represents the "top" of the adjacent pairs that might extend |
84d4ea48 |
156 | ** the run beginning at b. Usually, t addresses a pair |
157 | ** that compares with opposite sense from (b,b+1). |
158 | ** However, it may also address a singleton element at the end of list1, |
a0288114 |
159 | ** or it may be equal to "last", the first element beyond list1. |
84d4ea48 |
160 | ** |
161 | ** r addresses the Nth pair following b. If this would be beyond t, |
162 | ** we back it off to t. Only when r is less than t do we consider the |
163 | ** run long enough to consider checking. |
164 | ** |
165 | ** q addresses a pair such that the pairs at b through q already form a run. |
166 | ** Often, q will equal b, indicating we only are sure of the pair itself. |
167 | ** However, a search on the previous cycle may have revealed a longer run, |
168 | ** so q may be greater than b. |
169 | ** |
170 | ** p is used to work back from a candidate r, trying to reach q, |
171 | ** which would mean b through r would be a run. If we discover such a run, |
172 | ** we start q at r and try to push it further towards t. |
173 | ** If b through r is NOT a run, we detect the wrong order at (p-1,p). |
174 | ** In any event, after the check (if any), we have two main cases. |
175 | ** |
176 | ** 1) Short run. b <= q < p <= r <= t. |
177 | ** b through q is a run (perhaps trivial) |
178 | ** q through p are uninteresting pairs |
179 | ** p through r is a run |
180 | ** |
181 | ** 2) Long run. b < r <= q < t. |
182 | ** b through q is a run (of length >= 2 * PTHRESH) |
183 | ** |
184 | ** Note that degenerate cases are not only possible, but likely. |
185 | ** For example, if the pair following b compares with opposite sense, |
186 | ** then b == q < p == r == t. |
187 | */ |
188 | |
189 | |
957d8989 |
190 | static IV |
84d4ea48 |
191 | dynprep(pTHX_ gptr *list1, gptr *list2, size_t nmemb, SVCOMPARE_t cmp) |
192 | { |
957d8989 |
193 | I32 sense; |
84d4ea48 |
194 | register gptr *b, *p, *q, *t, *p2; |
195 | register gptr c, *last, *r; |
196 | gptr *savep; |
957d8989 |
197 | IV runs = 0; |
84d4ea48 |
198 | |
199 | b = list1; |
200 | last = PINDEX(b, nmemb); |
201 | sense = (cmp(aTHX_ *b, *(b+1)) > 0); |
202 | for (p2 = list2; b < last; ) { |
203 | /* We just started, or just reversed sense. |
204 | ** Set t at end of pairs with the prevailing sense. |
205 | */ |
206 | for (p = b+2, t = p; ++p < last; t = ++p) { |
207 | if ((cmp(aTHX_ *t, *p) > 0) != sense) break; |
208 | } |
209 | q = b; |
210 | /* Having laid out the playing field, look for long runs */ |
211 | do { |
212 | p = r = b + (2 * PTHRESH); |
213 | if (r >= t) p = r = t; /* too short to care about */ |
214 | else { |
215 | while (((cmp(aTHX_ *(p-1), *p) > 0) == sense) && |
216 | ((p -= 2) > q)); |
217 | if (p <= q) { |
218 | /* b through r is a (long) run. |
219 | ** Extend it as far as possible. |
220 | */ |
221 | p = q = r; |
222 | while (((p += 2) < t) && |
223 | ((cmp(aTHX_ *(p-1), *p) > 0) == sense)) q = p; |
224 | r = p = q + 2; /* no simple pairs, no after-run */ |
225 | } |
226 | } |
227 | if (q > b) { /* run of greater than 2 at b */ |
228 | savep = p; |
229 | p = q += 2; |
230 | /* pick up singleton, if possible */ |
231 | if ((p == t) && |
232 | ((t + 1) == last) && |
233 | ((cmp(aTHX_ *(p-1), *p) > 0) == sense)) |
234 | savep = r = p = q = last; |
957d8989 |
235 | p2 = NEXT(p2) = p2 + (p - b); ++runs; |
84d4ea48 |
236 | if (sense) while (b < --p) { |
237 | c = *b; |
238 | *b++ = *p; |
239 | *p = c; |
240 | } |
241 | p = savep; |
242 | } |
243 | while (q < p) { /* simple pairs */ |
957d8989 |
244 | p2 = NEXT(p2) = p2 + 2; ++runs; |
84d4ea48 |
245 | if (sense) { |
246 | c = *q++; |
247 | *(q-1) = *q; |
248 | *q++ = c; |
249 | } else q += 2; |
250 | } |
251 | if (((b = p) == t) && ((t+1) == last)) { |
957d8989 |
252 | NEXT(p2) = p2 + 1; ++runs; |
84d4ea48 |
253 | b++; |
254 | } |
255 | q = r; |
256 | } while (b < t); |
257 | sense = !sense; |
258 | } |
957d8989 |
259 | return runs; |
84d4ea48 |
260 | } |
261 | |
262 | |
3fe0b9a9 |
263 | /* The original merge sort, in use since 5.7, was as fast as, or faster than, |
957d8989 |
264 | * qsort on many platforms, but slower than qsort, conspicuously so, |
3fe0b9a9 |
265 | * on others. The most likely explanation was platform-specific |
957d8989 |
266 | * differences in cache sizes and relative speeds. |
267 | * |
268 | * The quicksort divide-and-conquer algorithm guarantees that, as the |
269 | * problem is subdivided into smaller and smaller parts, the parts |
270 | * fit into smaller (and faster) caches. So it doesn't matter how |
271 | * many levels of cache exist, quicksort will "find" them, and, |
272 | * as long as smaller is faster, take advanatge of them. |
273 | * |
3fe0b9a9 |
274 | * By contrast, consider how the original mergesort algorithm worked. |
957d8989 |
275 | * Suppose we have five runs (each typically of length 2 after dynprep). |
276 | * |
277 | * pass base aux |
278 | * 0 1 2 3 4 5 |
279 | * 1 12 34 5 |
280 | * 2 1234 5 |
281 | * 3 12345 |
282 | * 4 12345 |
283 | * |
284 | * Adjacent pairs are merged in "grand sweeps" through the input. |
285 | * This means, on pass 1, the records in runs 1 and 2 aren't revisited until |
286 | * runs 3 and 4 are merged and the runs from run 5 have been copied. |
287 | * The only cache that matters is one large enough to hold *all* the input. |
288 | * On some platforms, this may be many times slower than smaller caches. |
289 | * |
290 | * The following pseudo-code uses the same basic merge algorithm, |
291 | * but in a divide-and-conquer way. |
292 | * |
293 | * # merge $runs runs at offset $offset of list $list1 into $list2. |
294 | * # all unmerged runs ($runs == 1) originate in list $base. |
295 | * sub mgsort2 { |
296 | * my ($offset, $runs, $base, $list1, $list2) = @_; |
297 | * |
298 | * if ($runs == 1) { |
299 | * if ($list1 is $base) copy run to $list2 |
300 | * return offset of end of list (or copy) |
301 | * } else { |
302 | * $off2 = mgsort2($offset, $runs-($runs/2), $base, $list2, $list1) |
303 | * mgsort2($off2, $runs/2, $base, $list2, $list1) |
304 | * merge the adjacent runs at $offset of $list1 into $list2 |
305 | * return the offset of the end of the merged runs |
306 | * } |
307 | * } |
308 | * mgsort2(0, $runs, $base, $aux, $base); |
309 | * |
310 | * For our 5 runs, the tree of calls looks like |
311 | * |
312 | * 5 |
313 | * 3 2 |
314 | * 2 1 1 1 |
315 | * 1 1 |
316 | * |
317 | * 1 2 3 4 5 |
318 | * |
319 | * and the corresponding activity looks like |
320 | * |
321 | * copy runs 1 and 2 from base to aux |
322 | * merge runs 1 and 2 from aux to base |
323 | * (run 3 is where it belongs, no copy needed) |
324 | * merge runs 12 and 3 from base to aux |
325 | * (runs 4 and 5 are where they belong, no copy needed) |
326 | * merge runs 4 and 5 from base to aux |
327 | * merge runs 123 and 45 from aux to base |
328 | * |
329 | * Note that we merge runs 1 and 2 immediately after copying them, |
330 | * while they are still likely to be in fast cache. Similarly, |
331 | * run 3 is merged with run 12 while it still may be lingering in cache. |
332 | * This implementation should therefore enjoy much of the cache-friendly |
333 | * behavior that quicksort does. In addition, it does less copying |
334 | * than the original mergesort implementation (only runs 1 and 2 are copied) |
335 | * and the "balancing" of merges is better (merged runs comprise more nearly |
336 | * equal numbers of original runs). |
337 | * |
338 | * The actual cache-friendly implementation will use a pseudo-stack |
339 | * to avoid recursion, and will unroll processing of runs of length 2, |
340 | * but it is otherwise similar to the recursive implementation. |
957d8989 |
341 | */ |
342 | |
343 | typedef struct { |
344 | IV offset; /* offset of 1st of 2 runs at this level */ |
345 | IV runs; /* how many runs must be combined into 1 */ |
346 | } off_runs; /* pseudo-stack element */ |
347 | |
6c3fb703 |
348 | |
349 | static I32 |
350 | cmp_desc(pTHX_ gptr a, gptr b) |
351 | { |
352 | return -PL_sort_RealCmp(aTHX_ a, b); |
353 | } |
354 | |
957d8989 |
355 | STATIC void |
6c3fb703 |
356 | S_mergesortsv(pTHX_ gptr *base, size_t nmemb, SVCOMPARE_t cmp, U32 flags) |
957d8989 |
357 | { |
358 | IV i, run, runs, offset; |
359 | I32 sense, level; |
360 | int iwhich; |
361 | register gptr *f1, *f2, *t, *b, *p, *tp2, *l1, *l2, *q; |
362 | gptr *aux, *list1, *list2; |
363 | gptr *p1; |
364 | gptr small[SMALLSORT]; |
365 | gptr *which[3]; |
366 | off_runs stack[60], *stackp; |
a80036c6 |
367 | SVCOMPARE_t savecmp = 0; |
957d8989 |
368 | |
369 | if (nmemb <= 1) return; /* sorted trivially */ |
6c3fb703 |
370 | |
371 | if (flags) { |
372 | savecmp = PL_sort_RealCmp; /* Save current comparison routine, if any */ |
373 | PL_sort_RealCmp = cmp; /* Put comparison routine where cmp_desc can find it */ |
374 | cmp = cmp_desc; |
375 | } |
376 | |
957d8989 |
377 | if (nmemb <= SMALLSORT) aux = small; /* use stack for aux array */ |
a02a5408 |
378 | else { Newx(aux,nmemb,gptr); } /* allocate auxilliary array */ |
957d8989 |
379 | level = 0; |
380 | stackp = stack; |
381 | stackp->runs = dynprep(aTHX_ base, aux, nmemb, cmp); |
382 | stackp->offset = offset = 0; |
383 | which[0] = which[2] = base; |
384 | which[1] = aux; |
385 | for (;;) { |
386 | /* On levels where both runs have be constructed (stackp->runs == 0), |
387 | * merge them, and note the offset of their end, in case the offset |
388 | * is needed at the next level up. Hop up a level, and, |
389 | * as long as stackp->runs is 0, keep merging. |
390 | */ |
391 | if ((runs = stackp->runs) == 0) { |
392 | iwhich = level & 1; |
393 | list1 = which[iwhich]; /* area where runs are now */ |
394 | list2 = which[++iwhich]; /* area for merged runs */ |
395 | do { |
396 | offset = stackp->offset; |
397 | f1 = p1 = list1 + offset; /* start of first run */ |
398 | p = tp2 = list2 + offset; /* where merged run will go */ |
399 | t = NEXT(p); /* where first run ends */ |
400 | f2 = l1 = POTHER(t, list2, list1); /* ... on the other side */ |
401 | t = NEXT(t); /* where second runs ends */ |
402 | l2 = POTHER(t, list2, list1); /* ... on the other side */ |
403 | offset = PNELEM(list2, t); |
404 | while (f1 < l1 && f2 < l2) { |
405 | /* If head 1 is larger than head 2, find ALL the elements |
406 | ** in list 2 strictly less than head1, write them all, |
407 | ** then head 1. Then compare the new heads, and repeat, |
408 | ** until one or both lists are exhausted. |
409 | ** |
410 | ** In all comparisons (after establishing |
411 | ** which head to merge) the item to merge |
412 | ** (at pointer q) is the first operand of |
413 | ** the comparison. When we want to know |
a0288114 |
414 | ** if "q is strictly less than the other", |
957d8989 |
415 | ** we can't just do |
416 | ** cmp(q, other) < 0 |
417 | ** because stability demands that we treat equality |
418 | ** as high when q comes from l2, and as low when |
419 | ** q was from l1. So we ask the question by doing |
420 | ** cmp(q, other) <= sense |
421 | ** and make sense == 0 when equality should look low, |
422 | ** and -1 when equality should look high. |
423 | */ |
424 | |
425 | |
426 | if (cmp(aTHX_ *f1, *f2) <= 0) { |
427 | q = f2; b = f1; t = l1; |
428 | sense = -1; |
429 | } else { |
430 | q = f1; b = f2; t = l2; |
431 | sense = 0; |
432 | } |
433 | |
434 | |
435 | /* ramp up |
436 | ** |
437 | ** Leave t at something strictly |
438 | ** greater than q (or at the end of the list), |
439 | ** and b at something strictly less than q. |
440 | */ |
441 | for (i = 1, run = 0 ;;) { |
442 | if ((p = PINDEX(b, i)) >= t) { |
443 | /* off the end */ |
444 | if (((p = PINDEX(t, -1)) > b) && |
445 | (cmp(aTHX_ *q, *p) <= sense)) |
446 | t = p; |
447 | else b = p; |
448 | break; |
449 | } else if (cmp(aTHX_ *q, *p) <= sense) { |
450 | t = p; |
451 | break; |
452 | } else b = p; |
453 | if (++run >= RTHRESH) i += i; |
454 | } |
455 | |
456 | |
457 | /* q is known to follow b and must be inserted before t. |
458 | ** Increment b, so the range of possibilities is [b,t). |
459 | ** Round binary split down, to favor early appearance. |
460 | ** Adjust b and t until q belongs just before t. |
461 | */ |
462 | |
463 | b++; |
464 | while (b < t) { |
465 | p = PINDEX(b, (PNELEM(b, t) - 1) / 2); |
466 | if (cmp(aTHX_ *q, *p) <= sense) { |
467 | t = p; |
468 | } else b = p + 1; |
469 | } |
470 | |
471 | |
472 | /* Copy all the strictly low elements */ |
473 | |
474 | if (q == f1) { |
475 | FROMTOUPTO(f2, tp2, t); |
476 | *tp2++ = *f1++; |
477 | } else { |
478 | FROMTOUPTO(f1, tp2, t); |
479 | *tp2++ = *f2++; |
480 | } |
481 | } |
482 | |
483 | |
484 | /* Run out remaining list */ |
485 | if (f1 == l1) { |
486 | if (f2 < l2) FROMTOUPTO(f2, tp2, l2); |
487 | } else FROMTOUPTO(f1, tp2, l1); |
488 | p1 = NEXT(p1) = POTHER(tp2, list2, list1); |
489 | |
490 | if (--level == 0) goto done; |
491 | --stackp; |
492 | t = list1; list1 = list2; list2 = t; /* swap lists */ |
493 | } while ((runs = stackp->runs) == 0); |
494 | } |
495 | |
496 | |
497 | stackp->runs = 0; /* current run will finish level */ |
498 | /* While there are more than 2 runs remaining, |
499 | * turn them into exactly 2 runs (at the "other" level), |
500 | * each made up of approximately half the runs. |
501 | * Stack the second half for later processing, |
502 | * and set about producing the first half now. |
503 | */ |
504 | while (runs > 2) { |
505 | ++level; |
506 | ++stackp; |
507 | stackp->offset = offset; |
508 | runs -= stackp->runs = runs / 2; |
509 | } |
510 | /* We must construct a single run from 1 or 2 runs. |
511 | * All the original runs are in which[0] == base. |
512 | * The run we construct must end up in which[level&1]. |
513 | */ |
514 | iwhich = level & 1; |
515 | if (runs == 1) { |
516 | /* Constructing a single run from a single run. |
517 | * If it's where it belongs already, there's nothing to do. |
518 | * Otherwise, copy it to where it belongs. |
519 | * A run of 1 is either a singleton at level 0, |
520 | * or the second half of a split 3. In neither event |
521 | * is it necessary to set offset. It will be set by the merge |
522 | * that immediately follows. |
523 | */ |
524 | if (iwhich) { /* Belongs in aux, currently in base */ |
525 | f1 = b = PINDEX(base, offset); /* where list starts */ |
526 | f2 = PINDEX(aux, offset); /* where list goes */ |
527 | t = NEXT(f2); /* where list will end */ |
528 | offset = PNELEM(aux, t); /* offset thereof */ |
529 | t = PINDEX(base, offset); /* where it currently ends */ |
530 | FROMTOUPTO(f1, f2, t); /* copy */ |
531 | NEXT(b) = t; /* set up parallel pointer */ |
532 | } else if (level == 0) goto done; /* single run at level 0 */ |
533 | } else { |
534 | /* Constructing a single run from two runs. |
535 | * The merge code at the top will do that. |
536 | * We need only make sure the two runs are in the "other" array, |
537 | * so they'll end up in the correct array after the merge. |
538 | */ |
539 | ++level; |
540 | ++stackp; |
541 | stackp->offset = offset; |
542 | stackp->runs = 0; /* take care of both runs, trigger merge */ |
543 | if (!iwhich) { /* Merged runs belong in aux, copy 1st */ |
544 | f1 = b = PINDEX(base, offset); /* where first run starts */ |
545 | f2 = PINDEX(aux, offset); /* where it will be copied */ |
546 | t = NEXT(f2); /* where first run will end */ |
547 | offset = PNELEM(aux, t); /* offset thereof */ |
548 | p = PINDEX(base, offset); /* end of first run */ |
549 | t = NEXT(t); /* where second run will end */ |
550 | t = PINDEX(base, PNELEM(aux, t)); /* where it now ends */ |
551 | FROMTOUPTO(f1, f2, t); /* copy both runs */ |
552 | NEXT(b) = p; /* paralled pointer for 1st */ |
553 | NEXT(p) = t; /* ... and for second */ |
554 | } |
555 | } |
556 | } |
557 | done: |
558 | if (aux != small) Safefree(aux); /* free iff allocated */ |
6c3fb703 |
559 | if (flags) { |
560 | PL_sort_RealCmp = savecmp; /* Restore current comparison routine, if any */ |
561 | } |
957d8989 |
562 | return; |
563 | } |
564 | |
84d4ea48 |
565 | /* |
566 | * The quicksort implementation was derived from source code contributed |
567 | * by Tom Horsley. |
568 | * |
569 | * NOTE: this code was derived from Tom Horsley's qsort replacement |
570 | * and should not be confused with the original code. |
571 | */ |
572 | |
573 | /* Copyright (C) Tom Horsley, 1997. All rights reserved. |
574 | |
575 | Permission granted to distribute under the same terms as perl which are |
576 | (briefly): |
577 | |
578 | This program is free software; you can redistribute it and/or modify |
579 | it under the terms of either: |
580 | |
581 | a) the GNU General Public License as published by the Free |
582 | Software Foundation; either version 1, or (at your option) any |
583 | later version, or |
584 | |
585 | b) the "Artistic License" which comes with this Kit. |
586 | |
587 | Details on the perl license can be found in the perl source code which |
588 | may be located via the www.perl.com web page. |
589 | |
590 | This is the most wonderfulest possible qsort I can come up with (and |
591 | still be mostly portable) My (limited) tests indicate it consistently |
592 | does about 20% fewer calls to compare than does the qsort in the Visual |
593 | C++ library, other vendors may vary. |
594 | |
595 | Some of the ideas in here can be found in "Algorithms" by Sedgewick, |
596 | others I invented myself (or more likely re-invented since they seemed |
597 | pretty obvious once I watched the algorithm operate for a while). |
598 | |
599 | Most of this code was written while watching the Marlins sweep the Giants |
600 | in the 1997 National League Playoffs - no Braves fans allowed to use this |
601 | code (just kidding :-). |
602 | |
603 | I realize that if I wanted to be true to the perl tradition, the only |
604 | comment in this file would be something like: |
605 | |
606 | ...they shuffled back towards the rear of the line. 'No, not at the |
607 | rear!' the slave-driver shouted. 'Three files up. And stay there... |
608 | |
609 | However, I really needed to violate that tradition just so I could keep |
610 | track of what happens myself, not to mention some poor fool trying to |
611 | understand this years from now :-). |
612 | */ |
613 | |
614 | /* ********************************************************** Configuration */ |
615 | |
616 | #ifndef QSORT_ORDER_GUESS |
617 | #define QSORT_ORDER_GUESS 2 /* Select doubling version of the netBSD trick */ |
618 | #endif |
619 | |
620 | /* QSORT_MAX_STACK is the largest number of partitions that can be stacked up for |
621 | future processing - a good max upper bound is log base 2 of memory size |
622 | (32 on 32 bit machines, 64 on 64 bit machines, etc). In reality can |
623 | safely be smaller than that since the program is taking up some space and |
624 | most operating systems only let you grab some subset of contiguous |
625 | memory (not to mention that you are normally sorting data larger than |
626 | 1 byte element size :-). |
627 | */ |
628 | #ifndef QSORT_MAX_STACK |
629 | #define QSORT_MAX_STACK 32 |
630 | #endif |
631 | |
632 | /* QSORT_BREAK_EVEN is the size of the largest partition we should insertion sort. |
633 | Anything bigger and we use qsort. If you make this too small, the qsort |
634 | will probably break (or become less efficient), because it doesn't expect |
635 | the middle element of a partition to be the same as the right or left - |
636 | you have been warned). |
637 | */ |
638 | #ifndef QSORT_BREAK_EVEN |
639 | #define QSORT_BREAK_EVEN 6 |
640 | #endif |
641 | |
4eb872f6 |
642 | /* QSORT_PLAY_SAFE is the size of the largest partition we're willing |
643 | to go quadratic on. We innoculate larger partitions against |
644 | quadratic behavior by shuffling them before sorting. This is not |
645 | an absolute guarantee of non-quadratic behavior, but it would take |
646 | staggeringly bad luck to pick extreme elements as the pivot |
647 | from randomized data. |
648 | */ |
649 | #ifndef QSORT_PLAY_SAFE |
650 | #define QSORT_PLAY_SAFE 255 |
651 | #endif |
652 | |
84d4ea48 |
653 | /* ************************************************************* Data Types */ |
654 | |
655 | /* hold left and right index values of a partition waiting to be sorted (the |
656 | partition includes both left and right - right is NOT one past the end or |
657 | anything like that). |
658 | */ |
659 | struct partition_stack_entry { |
660 | int left; |
661 | int right; |
662 | #ifdef QSORT_ORDER_GUESS |
663 | int qsort_break_even; |
664 | #endif |
665 | }; |
666 | |
667 | /* ******************************************************* Shorthand Macros */ |
668 | |
669 | /* Note that these macros will be used from inside the qsort function where |
670 | we happen to know that the variable 'elt_size' contains the size of an |
671 | array element and the variable 'temp' points to enough space to hold a |
672 | temp element and the variable 'array' points to the array being sorted |
673 | and 'compare' is the pointer to the compare routine. |
674 | |
675 | Also note that there are very many highly architecture specific ways |
676 | these might be sped up, but this is simply the most generally portable |
677 | code I could think of. |
678 | */ |
679 | |
680 | /* Return < 0 == 0 or > 0 as the value of elt1 is < elt2, == elt2, > elt2 |
681 | */ |
682 | #define qsort_cmp(elt1, elt2) \ |
683 | ((*compare)(aTHX_ array[elt1], array[elt2])) |
684 | |
685 | #ifdef QSORT_ORDER_GUESS |
686 | #define QSORT_NOTICE_SWAP swapped++; |
687 | #else |
688 | #define QSORT_NOTICE_SWAP |
689 | #endif |
690 | |
691 | /* swaps contents of array elements elt1, elt2. |
692 | */ |
693 | #define qsort_swap(elt1, elt2) \ |
694 | STMT_START { \ |
695 | QSORT_NOTICE_SWAP \ |
696 | temp = array[elt1]; \ |
697 | array[elt1] = array[elt2]; \ |
698 | array[elt2] = temp; \ |
699 | } STMT_END |
700 | |
701 | /* rotate contents of elt1, elt2, elt3 such that elt1 gets elt2, elt2 gets |
702 | elt3 and elt3 gets elt1. |
703 | */ |
704 | #define qsort_rotate(elt1, elt2, elt3) \ |
705 | STMT_START { \ |
706 | QSORT_NOTICE_SWAP \ |
707 | temp = array[elt1]; \ |
708 | array[elt1] = array[elt2]; \ |
709 | array[elt2] = array[elt3]; \ |
710 | array[elt3] = temp; \ |
711 | } STMT_END |
712 | |
713 | /* ************************************************************ Debug stuff */ |
714 | |
715 | #ifdef QSORT_DEBUG |
716 | |
717 | static void |
718 | break_here() |
719 | { |
720 | return; /* good place to set a breakpoint */ |
721 | } |
722 | |
723 | #define qsort_assert(t) (void)( (t) || (break_here(), 0) ) |
724 | |
725 | static void |
726 | doqsort_all_asserts( |
727 | void * array, |
728 | size_t num_elts, |
729 | size_t elt_size, |
730 | int (*compare)(const void * elt1, const void * elt2), |
731 | int pc_left, int pc_right, int u_left, int u_right) |
732 | { |
733 | int i; |
734 | |
735 | qsort_assert(pc_left <= pc_right); |
736 | qsort_assert(u_right < pc_left); |
737 | qsort_assert(pc_right < u_left); |
738 | for (i = u_right + 1; i < pc_left; ++i) { |
739 | qsort_assert(qsort_cmp(i, pc_left) < 0); |
740 | } |
741 | for (i = pc_left; i < pc_right; ++i) { |
742 | qsort_assert(qsort_cmp(i, pc_right) == 0); |
743 | } |
744 | for (i = pc_right + 1; i < u_left; ++i) { |
745 | qsort_assert(qsort_cmp(pc_right, i) < 0); |
746 | } |
747 | } |
748 | |
749 | #define qsort_all_asserts(PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT) \ |
750 | doqsort_all_asserts(array, num_elts, elt_size, compare, \ |
751 | PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT) |
752 | |
753 | #else |
754 | |
755 | #define qsort_assert(t) ((void)0) |
756 | |
757 | #define qsort_all_asserts(PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT) ((void)0) |
758 | |
759 | #endif |
760 | |
761 | /* ****************************************************************** qsort */ |
762 | |
763 | STATIC void /* the standard unstable (u) quicksort (qsort) */ |
764 | S_qsortsvu(pTHX_ SV ** array, size_t num_elts, SVCOMPARE_t compare) |
765 | { |
766 | register SV * temp; |
767 | |
768 | struct partition_stack_entry partition_stack[QSORT_MAX_STACK]; |
769 | int next_stack_entry = 0; |
770 | |
771 | int part_left; |
772 | int part_right; |
773 | #ifdef QSORT_ORDER_GUESS |
774 | int qsort_break_even; |
775 | int swapped; |
776 | #endif |
777 | |
778 | /* Make sure we actually have work to do. |
779 | */ |
780 | if (num_elts <= 1) { |
781 | return; |
782 | } |
783 | |
4eb872f6 |
784 | /* Innoculate large partitions against quadratic behavior */ |
785 | if (num_elts > QSORT_PLAY_SAFE) { |
901017d6 |
786 | register size_t n; |
787 | register SV ** const q = array; |
788 | for (n = num_elts; n > 1; ) { |
789 | register const size_t j = (size_t)(n-- * Drand01()); |
4eb872f6 |
790 | temp = q[j]; |
791 | q[j] = q[n]; |
792 | q[n] = temp; |
793 | } |
794 | } |
795 | |
84d4ea48 |
796 | /* Setup the initial partition definition and fall into the sorting loop |
797 | */ |
798 | part_left = 0; |
799 | part_right = (int)(num_elts - 1); |
800 | #ifdef QSORT_ORDER_GUESS |
801 | qsort_break_even = QSORT_BREAK_EVEN; |
802 | #else |
803 | #define qsort_break_even QSORT_BREAK_EVEN |
804 | #endif |
805 | for ( ; ; ) { |
806 | if ((part_right - part_left) >= qsort_break_even) { |
807 | /* OK, this is gonna get hairy, so lets try to document all the |
808 | concepts and abbreviations and variables and what they keep |
809 | track of: |
810 | |
811 | pc: pivot chunk - the set of array elements we accumulate in the |
812 | middle of the partition, all equal in value to the original |
813 | pivot element selected. The pc is defined by: |
814 | |
815 | pc_left - the leftmost array index of the pc |
816 | pc_right - the rightmost array index of the pc |
817 | |
818 | we start with pc_left == pc_right and only one element |
819 | in the pivot chunk (but it can grow during the scan). |
820 | |
821 | u: uncompared elements - the set of elements in the partition |
822 | we have not yet compared to the pivot value. There are two |
823 | uncompared sets during the scan - one to the left of the pc |
824 | and one to the right. |
825 | |
826 | u_right - the rightmost index of the left side's uncompared set |
827 | u_left - the leftmost index of the right side's uncompared set |
828 | |
829 | The leftmost index of the left sides's uncompared set |
830 | doesn't need its own variable because it is always defined |
831 | by the leftmost edge of the whole partition (part_left). The |
832 | same goes for the rightmost edge of the right partition |
833 | (part_right). |
834 | |
835 | We know there are no uncompared elements on the left once we |
836 | get u_right < part_left and no uncompared elements on the |
837 | right once u_left > part_right. When both these conditions |
838 | are met, we have completed the scan of the partition. |
839 | |
840 | Any elements which are between the pivot chunk and the |
841 | uncompared elements should be less than the pivot value on |
842 | the left side and greater than the pivot value on the right |
843 | side (in fact, the goal of the whole algorithm is to arrange |
844 | for that to be true and make the groups of less-than and |
845 | greater-then elements into new partitions to sort again). |
846 | |
847 | As you marvel at the complexity of the code and wonder why it |
848 | has to be so confusing. Consider some of the things this level |
849 | of confusion brings: |
850 | |
851 | Once I do a compare, I squeeze every ounce of juice out of it. I |
852 | never do compare calls I don't have to do, and I certainly never |
853 | do redundant calls. |
854 | |
855 | I also never swap any elements unless I can prove there is a |
856 | good reason. Many sort algorithms will swap a known value with |
857 | an uncompared value just to get things in the right place (or |
858 | avoid complexity :-), but that uncompared value, once it gets |
859 | compared, may then have to be swapped again. A lot of the |
860 | complexity of this code is due to the fact that it never swaps |
861 | anything except compared values, and it only swaps them when the |
862 | compare shows they are out of position. |
863 | */ |
864 | int pc_left, pc_right; |
865 | int u_right, u_left; |
866 | |
867 | int s; |
868 | |
869 | pc_left = ((part_left + part_right) / 2); |
870 | pc_right = pc_left; |
871 | u_right = pc_left - 1; |
872 | u_left = pc_right + 1; |
873 | |
874 | /* Qsort works best when the pivot value is also the median value |
875 | in the partition (unfortunately you can't find the median value |
876 | without first sorting :-), so to give the algorithm a helping |
877 | hand, we pick 3 elements and sort them and use the median value |
878 | of that tiny set as the pivot value. |
879 | |
880 | Some versions of qsort like to use the left middle and right as |
881 | the 3 elements to sort so they can insure the ends of the |
882 | partition will contain values which will stop the scan in the |
883 | compare loop, but when you have to call an arbitrarily complex |
884 | routine to do a compare, its really better to just keep track of |
885 | array index values to know when you hit the edge of the |
886 | partition and avoid the extra compare. An even better reason to |
887 | avoid using a compare call is the fact that you can drop off the |
888 | edge of the array if someone foolishly provides you with an |
889 | unstable compare function that doesn't always provide consistent |
890 | results. |
891 | |
892 | So, since it is simpler for us to compare the three adjacent |
893 | elements in the middle of the partition, those are the ones we |
894 | pick here (conveniently pointed at by u_right, pc_left, and |
895 | u_left). The values of the left, center, and right elements |
896 | are refered to as l c and r in the following comments. |
897 | */ |
898 | |
899 | #ifdef QSORT_ORDER_GUESS |
900 | swapped = 0; |
901 | #endif |
902 | s = qsort_cmp(u_right, pc_left); |
903 | if (s < 0) { |
904 | /* l < c */ |
905 | s = qsort_cmp(pc_left, u_left); |
906 | /* if l < c, c < r - already in order - nothing to do */ |
907 | if (s == 0) { |
908 | /* l < c, c == r - already in order, pc grows */ |
909 | ++pc_right; |
910 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
911 | } else if (s > 0) { |
912 | /* l < c, c > r - need to know more */ |
913 | s = qsort_cmp(u_right, u_left); |
914 | if (s < 0) { |
915 | /* l < c, c > r, l < r - swap c & r to get ordered */ |
916 | qsort_swap(pc_left, u_left); |
917 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
918 | } else if (s == 0) { |
919 | /* l < c, c > r, l == r - swap c&r, grow pc */ |
920 | qsort_swap(pc_left, u_left); |
921 | --pc_left; |
922 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
923 | } else { |
924 | /* l < c, c > r, l > r - make lcr into rlc to get ordered */ |
925 | qsort_rotate(pc_left, u_right, u_left); |
926 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
927 | } |
928 | } |
929 | } else if (s == 0) { |
930 | /* l == c */ |
931 | s = qsort_cmp(pc_left, u_left); |
932 | if (s < 0) { |
933 | /* l == c, c < r - already in order, grow pc */ |
934 | --pc_left; |
935 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
936 | } else if (s == 0) { |
937 | /* l == c, c == r - already in order, grow pc both ways */ |
938 | --pc_left; |
939 | ++pc_right; |
940 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
941 | } else { |
942 | /* l == c, c > r - swap l & r, grow pc */ |
943 | qsort_swap(u_right, u_left); |
944 | ++pc_right; |
945 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
946 | } |
947 | } else { |
948 | /* l > c */ |
949 | s = qsort_cmp(pc_left, u_left); |
950 | if (s < 0) { |
951 | /* l > c, c < r - need to know more */ |
952 | s = qsort_cmp(u_right, u_left); |
953 | if (s < 0) { |
954 | /* l > c, c < r, l < r - swap l & c to get ordered */ |
955 | qsort_swap(u_right, pc_left); |
956 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
957 | } else if (s == 0) { |
958 | /* l > c, c < r, l == r - swap l & c, grow pc */ |
959 | qsort_swap(u_right, pc_left); |
960 | ++pc_right; |
961 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
962 | } else { |
963 | /* l > c, c < r, l > r - rotate lcr into crl to order */ |
964 | qsort_rotate(u_right, pc_left, u_left); |
965 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
966 | } |
967 | } else if (s == 0) { |
968 | /* l > c, c == r - swap ends, grow pc */ |
969 | qsort_swap(u_right, u_left); |
970 | --pc_left; |
971 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
972 | } else { |
973 | /* l > c, c > r - swap ends to get in order */ |
974 | qsort_swap(u_right, u_left); |
975 | qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1); |
976 | } |
977 | } |
978 | /* We now know the 3 middle elements have been compared and |
979 | arranged in the desired order, so we can shrink the uncompared |
980 | sets on both sides |
981 | */ |
982 | --u_right; |
983 | ++u_left; |
984 | qsort_all_asserts(pc_left, pc_right, u_left, u_right); |
985 | |
986 | /* The above massive nested if was the simple part :-). We now have |
987 | the middle 3 elements ordered and we need to scan through the |
988 | uncompared sets on either side, swapping elements that are on |
989 | the wrong side or simply shuffling equal elements around to get |
990 | all equal elements into the pivot chunk. |
991 | */ |
992 | |
993 | for ( ; ; ) { |
994 | int still_work_on_left; |
995 | int still_work_on_right; |
996 | |
997 | /* Scan the uncompared values on the left. If I find a value |
998 | equal to the pivot value, move it over so it is adjacent to |
999 | the pivot chunk and expand the pivot chunk. If I find a value |
1000 | less than the pivot value, then just leave it - its already |
1001 | on the correct side of the partition. If I find a greater |
1002 | value, then stop the scan. |
1003 | */ |
1004 | while ((still_work_on_left = (u_right >= part_left))) { |
1005 | s = qsort_cmp(u_right, pc_left); |
1006 | if (s < 0) { |
1007 | --u_right; |
1008 | } else if (s == 0) { |
1009 | --pc_left; |
1010 | if (pc_left != u_right) { |
1011 | qsort_swap(u_right, pc_left); |
1012 | } |
1013 | --u_right; |
1014 | } else { |
1015 | break; |
1016 | } |
1017 | qsort_assert(u_right < pc_left); |
1018 | qsort_assert(pc_left <= pc_right); |
1019 | qsort_assert(qsort_cmp(u_right + 1, pc_left) <= 0); |
1020 | qsort_assert(qsort_cmp(pc_left, pc_right) == 0); |
1021 | } |
1022 | |
1023 | /* Do a mirror image scan of uncompared values on the right |
1024 | */ |
1025 | while ((still_work_on_right = (u_left <= part_right))) { |
1026 | s = qsort_cmp(pc_right, u_left); |
1027 | if (s < 0) { |
1028 | ++u_left; |
1029 | } else if (s == 0) { |
1030 | ++pc_right; |
1031 | if (pc_right != u_left) { |
1032 | qsort_swap(pc_right, u_left); |
1033 | } |
1034 | ++u_left; |
1035 | } else { |
1036 | break; |
1037 | } |
1038 | qsort_assert(u_left > pc_right); |
1039 | qsort_assert(pc_left <= pc_right); |
1040 | qsort_assert(qsort_cmp(pc_right, u_left - 1) <= 0); |
1041 | qsort_assert(qsort_cmp(pc_left, pc_right) == 0); |
1042 | } |
1043 | |
1044 | if (still_work_on_left) { |
1045 | /* I know I have a value on the left side which needs to be |
1046 | on the right side, but I need to know more to decide |
1047 | exactly the best thing to do with it. |
1048 | */ |
1049 | if (still_work_on_right) { |
1050 | /* I know I have values on both side which are out of |
1051 | position. This is a big win because I kill two birds |
1052 | with one swap (so to speak). I can advance the |
1053 | uncompared pointers on both sides after swapping both |
1054 | of them into the right place. |
1055 | */ |
1056 | qsort_swap(u_right, u_left); |
1057 | --u_right; |
1058 | ++u_left; |
1059 | qsort_all_asserts(pc_left, pc_right, u_left, u_right); |
1060 | } else { |
1061 | /* I have an out of position value on the left, but the |
1062 | right is fully scanned, so I "slide" the pivot chunk |
1063 | and any less-than values left one to make room for the |
1064 | greater value over on the right. If the out of position |
1065 | value is immediately adjacent to the pivot chunk (there |
1066 | are no less-than values), I can do that with a swap, |
1067 | otherwise, I have to rotate one of the less than values |
1068 | into the former position of the out of position value |
1069 | and the right end of the pivot chunk into the left end |
1070 | (got all that?). |
1071 | */ |
1072 | --pc_left; |
1073 | if (pc_left == u_right) { |
1074 | qsort_swap(u_right, pc_right); |
1075 | qsort_all_asserts(pc_left, pc_right-1, u_left, u_right-1); |
1076 | } else { |
1077 | qsort_rotate(u_right, pc_left, pc_right); |
1078 | qsort_all_asserts(pc_left, pc_right-1, u_left, u_right-1); |
1079 | } |
1080 | --pc_right; |
1081 | --u_right; |
1082 | } |
1083 | } else if (still_work_on_right) { |
1084 | /* Mirror image of complex case above: I have an out of |
1085 | position value on the right, but the left is fully |
1086 | scanned, so I need to shuffle things around to make room |
1087 | for the right value on the left. |
1088 | */ |
1089 | ++pc_right; |
1090 | if (pc_right == u_left) { |
1091 | qsort_swap(u_left, pc_left); |
1092 | qsort_all_asserts(pc_left+1, pc_right, u_left+1, u_right); |
1093 | } else { |
1094 | qsort_rotate(pc_right, pc_left, u_left); |
1095 | qsort_all_asserts(pc_left+1, pc_right, u_left+1, u_right); |
1096 | } |
1097 | ++pc_left; |
1098 | ++u_left; |
1099 | } else { |
1100 | /* No more scanning required on either side of partition, |
1101 | break out of loop and figure out next set of partitions |
1102 | */ |
1103 | break; |
1104 | } |
1105 | } |
1106 | |
1107 | /* The elements in the pivot chunk are now in the right place. They |
1108 | will never move or be compared again. All I have to do is decide |
1109 | what to do with the stuff to the left and right of the pivot |
1110 | chunk. |
1111 | |
1112 | Notes on the QSORT_ORDER_GUESS ifdef code: |
1113 | |
1114 | 1. If I just built these partitions without swapping any (or |
1115 | very many) elements, there is a chance that the elements are |
1116 | already ordered properly (being properly ordered will |
1117 | certainly result in no swapping, but the converse can't be |
1118 | proved :-). |
1119 | |
1120 | 2. A (properly written) insertion sort will run faster on |
1121 | already ordered data than qsort will. |
1122 | |
1123 | 3. Perhaps there is some way to make a good guess about |
1124 | switching to an insertion sort earlier than partition size 6 |
1125 | (for instance - we could save the partition size on the stack |
1126 | and increase the size each time we find we didn't swap, thus |
1127 | switching to insertion sort earlier for partitions with a |
1128 | history of not swapping). |
1129 | |
1130 | 4. Naturally, if I just switch right away, it will make |
1131 | artificial benchmarks with pure ascending (or descending) |
1132 | data look really good, but is that a good reason in general? |
1133 | Hard to say... |
1134 | */ |
1135 | |
1136 | #ifdef QSORT_ORDER_GUESS |
1137 | if (swapped < 3) { |
1138 | #if QSORT_ORDER_GUESS == 1 |
1139 | qsort_break_even = (part_right - part_left) + 1; |
1140 | #endif |
1141 | #if QSORT_ORDER_GUESS == 2 |
1142 | qsort_break_even *= 2; |
1143 | #endif |
1144 | #if QSORT_ORDER_GUESS == 3 |
901017d6 |
1145 | const int prev_break = qsort_break_even; |
84d4ea48 |
1146 | qsort_break_even *= qsort_break_even; |
1147 | if (qsort_break_even < prev_break) { |
1148 | qsort_break_even = (part_right - part_left) + 1; |
1149 | } |
1150 | #endif |
1151 | } else { |
1152 | qsort_break_even = QSORT_BREAK_EVEN; |
1153 | } |
1154 | #endif |
1155 | |
1156 | if (part_left < pc_left) { |
1157 | /* There are elements on the left which need more processing. |
1158 | Check the right as well before deciding what to do. |
1159 | */ |
1160 | if (pc_right < part_right) { |
1161 | /* We have two partitions to be sorted. Stack the biggest one |
1162 | and process the smallest one on the next iteration. This |
1163 | minimizes the stack height by insuring that any additional |
1164 | stack entries must come from the smallest partition which |
1165 | (because it is smallest) will have the fewest |
1166 | opportunities to generate additional stack entries. |
1167 | */ |
1168 | if ((part_right - pc_right) > (pc_left - part_left)) { |
1169 | /* stack the right partition, process the left */ |
1170 | partition_stack[next_stack_entry].left = pc_right + 1; |
1171 | partition_stack[next_stack_entry].right = part_right; |
1172 | #ifdef QSORT_ORDER_GUESS |
1173 | partition_stack[next_stack_entry].qsort_break_even = qsort_break_even; |
1174 | #endif |
1175 | part_right = pc_left - 1; |
1176 | } else { |
1177 | /* stack the left partition, process the right */ |
1178 | partition_stack[next_stack_entry].left = part_left; |
1179 | partition_stack[next_stack_entry].right = pc_left - 1; |
1180 | #ifdef QSORT_ORDER_GUESS |
1181 | partition_stack[next_stack_entry].qsort_break_even = qsort_break_even; |
1182 | #endif |
1183 | part_left = pc_right + 1; |
1184 | } |
1185 | qsort_assert(next_stack_entry < QSORT_MAX_STACK); |
1186 | ++next_stack_entry; |
1187 | } else { |
1188 | /* The elements on the left are the only remaining elements |
1189 | that need sorting, arrange for them to be processed as the |
1190 | next partition. |
1191 | */ |
1192 | part_right = pc_left - 1; |
1193 | } |
1194 | } else if (pc_right < part_right) { |
1195 | /* There is only one chunk on the right to be sorted, make it |
1196 | the new partition and loop back around. |
1197 | */ |
1198 | part_left = pc_right + 1; |
1199 | } else { |
1200 | /* This whole partition wound up in the pivot chunk, so |
1201 | we need to get a new partition off the stack. |
1202 | */ |
1203 | if (next_stack_entry == 0) { |
1204 | /* the stack is empty - we are done */ |
1205 | break; |
1206 | } |
1207 | --next_stack_entry; |
1208 | part_left = partition_stack[next_stack_entry].left; |
1209 | part_right = partition_stack[next_stack_entry].right; |
1210 | #ifdef QSORT_ORDER_GUESS |
1211 | qsort_break_even = partition_stack[next_stack_entry].qsort_break_even; |
1212 | #endif |
1213 | } |
1214 | } else { |
1215 | /* This partition is too small to fool with qsort complexity, just |
1216 | do an ordinary insertion sort to minimize overhead. |
1217 | */ |
1218 | int i; |
1219 | /* Assume 1st element is in right place already, and start checking |
1220 | at 2nd element to see where it should be inserted. |
1221 | */ |
1222 | for (i = part_left + 1; i <= part_right; ++i) { |
1223 | int j; |
1224 | /* Scan (backwards - just in case 'i' is already in right place) |
1225 | through the elements already sorted to see if the ith element |
1226 | belongs ahead of one of them. |
1227 | */ |
1228 | for (j = i - 1; j >= part_left; --j) { |
1229 | if (qsort_cmp(i, j) >= 0) { |
1230 | /* i belongs right after j |
1231 | */ |
1232 | break; |
1233 | } |
1234 | } |
1235 | ++j; |
1236 | if (j != i) { |
1237 | /* Looks like we really need to move some things |
1238 | */ |
1239 | int k; |
1240 | temp = array[i]; |
1241 | for (k = i - 1; k >= j; --k) |
1242 | array[k + 1] = array[k]; |
1243 | array[j] = temp; |
1244 | } |
1245 | } |
1246 | |
1247 | /* That partition is now sorted, grab the next one, or get out |
1248 | of the loop if there aren't any more. |
1249 | */ |
1250 | |
1251 | if (next_stack_entry == 0) { |
1252 | /* the stack is empty - we are done */ |
1253 | break; |
1254 | } |
1255 | --next_stack_entry; |
1256 | part_left = partition_stack[next_stack_entry].left; |
1257 | part_right = partition_stack[next_stack_entry].right; |
1258 | #ifdef QSORT_ORDER_GUESS |
1259 | qsort_break_even = partition_stack[next_stack_entry].qsort_break_even; |
1260 | #endif |
1261 | } |
1262 | } |
1263 | |
1264 | /* Believe it or not, the array is sorted at this point! */ |
1265 | } |
1266 | |
84d4ea48 |
1267 | /* Stabilize what is, presumably, an otherwise unstable sort method. |
1268 | * We do that by allocating (or having on hand) an array of pointers |
1269 | * that is the same size as the original array of elements to be sorted. |
1270 | * We initialize this parallel array with the addresses of the original |
1271 | * array elements. This indirection can make you crazy. |
1272 | * Some pictures can help. After initializing, we have |
1273 | * |
1274 | * indir list1 |
1275 | * +----+ +----+ |
1276 | * | | --------------> | | ------> first element to be sorted |
1277 | * +----+ +----+ |
1278 | * | | --------------> | | ------> second element to be sorted |
1279 | * +----+ +----+ |
1280 | * | | --------------> | | ------> third element to be sorted |
1281 | * +----+ +----+ |
1282 | * ... |
1283 | * +----+ +----+ |
1284 | * | | --------------> | | ------> n-1st element to be sorted |
1285 | * +----+ +----+ |
1286 | * | | --------------> | | ------> n-th element to be sorted |
1287 | * +----+ +----+ |
1288 | * |
1289 | * During the sort phase, we leave the elements of list1 where they are, |
1290 | * and sort the pointers in the indirect array in the same order determined |
1291 | * by the original comparison routine on the elements pointed to. |
1292 | * Because we don't move the elements of list1 around through |
1293 | * this phase, we can break ties on elements that compare equal |
1294 | * using their address in the list1 array, ensuring stabilty. |
1295 | * This leaves us with something looking like |
1296 | * |
1297 | * indir list1 |
1298 | * +----+ +----+ |
1299 | * | | --+ +---> | | ------> first element to be sorted |
1300 | * +----+ | | +----+ |
1301 | * | | --|-------|---> | | ------> second element to be sorted |
1302 | * +----+ | | +----+ |
1303 | * | | --|-------+ +-> | | ------> third element to be sorted |
1304 | * +----+ | | +----+ |
1305 | * ... |
1306 | * +----+ | | | | +----+ |
1307 | * | | ---|-+ | +--> | | ------> n-1st element to be sorted |
1308 | * +----+ | | +----+ |
1309 | * | | ---+ +----> | | ------> n-th element to be sorted |
1310 | * +----+ +----+ |
1311 | * |
1312 | * where the i-th element of the indirect array points to the element |
1313 | * that should be i-th in the sorted array. After the sort phase, |
1314 | * we have to put the elements of list1 into the places |
1315 | * dictated by the indirect array. |
1316 | */ |
1317 | |
84d4ea48 |
1318 | |
1319 | static I32 |
1320 | cmpindir(pTHX_ gptr a, gptr b) |
1321 | { |
1322 | I32 sense; |
901017d6 |
1323 | gptr * const ap = (gptr *)a; |
1324 | gptr * const bp = (gptr *)b; |
84d4ea48 |
1325 | |
147f47de |
1326 | if ((sense = PL_sort_RealCmp(aTHX_ *ap, *bp)) == 0) |
84d4ea48 |
1327 | sense = (ap > bp) ? 1 : ((ap < bp) ? -1 : 0); |
1328 | return sense; |
1329 | } |
1330 | |
6c3fb703 |
1331 | static I32 |
1332 | cmpindir_desc(pTHX_ gptr a, gptr b) |
1333 | { |
1334 | I32 sense; |
901017d6 |
1335 | gptr * const ap = (gptr *)a; |
1336 | gptr * const bp = (gptr *)b; |
6c3fb703 |
1337 | |
1338 | /* Reverse the default */ |
1339 | if ((sense = PL_sort_RealCmp(aTHX_ *ap, *bp))) |
1340 | return -sense; |
1341 | /* But don't reverse the stability test. */ |
1342 | return (ap > bp) ? 1 : ((ap < bp) ? -1 : 0); |
1343 | |
1344 | } |
1345 | |
84d4ea48 |
1346 | STATIC void |
6c3fb703 |
1347 | S_qsortsv(pTHX_ gptr *list1, size_t nmemb, SVCOMPARE_t cmp, U32 flags) |
84d4ea48 |
1348 | { |
84d4ea48 |
1349 | |
5fe61d93 |
1350 | dSORTHINTS; |
1351 | |
1352 | if (SORTHINTS & 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; |
a02a5408 |
1361 | else { Newx(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; |
5fe61d93 |
1445 | dSORTHINTS; |
1446 | const I32 hints = SORTHINTS; |
78210658 |
1447 | if (hints & HINT_SORT_QUICKSORT) { |
1448 | sortsvp = S_qsortsv; |
1449 | } |
1450 | else { |
1451 | /* The default as of 5.8.0 is mergesort */ |
1452 | sortsvp = S_mergesortsv; |
84d4ea48 |
1453 | } |
4eb872f6 |
1454 | |
6c3fb703 |
1455 | sortsvp(aTHX_ array, nmemb, cmp, 0); |
1456 | } |
1457 | |
1458 | |
b7787f18 |
1459 | static void |
6c3fb703 |
1460 | S_sortsv_desc(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp) |
1461 | { |
1462 | void (*sortsvp)(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp, U32 flags) |
1463 | = S_mergesortsv; |
5fe61d93 |
1464 | dSORTHINTS; |
1465 | const I32 hints = SORTHINTS; |
6c3fb703 |
1466 | if (hints & HINT_SORT_QUICKSORT) { |
1467 | sortsvp = S_qsortsv; |
1468 | } |
1469 | else { |
1470 | /* The default as of 5.8.0 is mergesort */ |
1471 | sortsvp = S_mergesortsv; |
1472 | } |
1473 | |
1474 | sortsvp(aTHX_ array, nmemb, cmp, 1); |
84d4ea48 |
1475 | } |
1476 | |
4d562308 |
1477 | #define SvNSIOK(sv) ((SvFLAGS(sv) & SVf_NOK) || ((SvFLAGS(sv) & (SVf_IOK|SVf_IVisUV)) == SVf_IOK)) |
1478 | #define SvSIOK(sv) ((SvFLAGS(sv) & (SVf_IOK|SVf_IVisUV)) == SVf_IOK) |
1479 | #define SvNSIV(sv) ( SvNOK(sv) ? SvNVX(sv) : ( SvSIOK(sv) ? SvIVX(sv) : sv_2nv(sv) ) ) |
1480 | |
84d4ea48 |
1481 | PP(pp_sort) |
1482 | { |
27da23d5 |
1483 | dVAR; dSP; dMARK; dORIGMARK; |
fe1bc4cf |
1484 | register SV **p1 = ORIGMARK+1, **p2; |
1485 | register I32 max, i; |
1486 | AV* av = Nullav; |
84d4ea48 |
1487 | HV *stash; |
1488 | GV *gv; |
1489 | CV *cv = 0; |
1490 | I32 gimme = GIMME; |
1491 | OP* nextop = PL_op->op_next; |
1492 | I32 overloading = 0; |
1493 | bool hasargs = FALSE; |
1494 | I32 is_xsub = 0; |
fe1bc4cf |
1495 | I32 sorting_av = 0; |
901017d6 |
1496 | const U8 priv = PL_op->op_private; |
1497 | const U8 flags = PL_op->op_flags; |
6c3fb703 |
1498 | void (*sortsvp)(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp) |
1499 | = Perl_sortsv; |
4d562308 |
1500 | I32 all_SIVs = 1; |
84d4ea48 |
1501 | |
1502 | if (gimme != G_ARRAY) { |
1503 | SP = MARK; |
1504 | RETPUSHUNDEF; |
1505 | } |
1506 | |
1507 | ENTER; |
1508 | SAVEVPTR(PL_sortcop); |
471178c0 |
1509 | if (flags & OPf_STACKED) { |
1510 | if (flags & OPf_SPECIAL) { |
84d4ea48 |
1511 | OP *kid = cLISTOP->op_first->op_sibling; /* pass pushmark */ |
1512 | kid = kUNOP->op_first; /* pass rv2gv */ |
1513 | kid = kUNOP->op_first; /* pass leave */ |
1514 | PL_sortcop = kid->op_next; |
1515 | stash = CopSTASH(PL_curcop); |
1516 | } |
1517 | else { |
1518 | cv = sv_2cv(*++MARK, &stash, &gv, 0); |
1519 | if (cv && SvPOK(cv)) { |
349d4f2f |
1520 | const char *proto = SvPV_nolen_const((SV*)cv); |
84d4ea48 |
1521 | if (proto && strEQ(proto, "$$")) { |
1522 | hasargs = TRUE; |
1523 | } |
1524 | } |
1525 | if (!(cv && CvROOT(cv))) { |
1526 | if (cv && CvXSUB(cv)) { |
1527 | is_xsub = 1; |
1528 | } |
1529 | else if (gv) { |
1530 | SV *tmpstr = sv_newmortal(); |
1531 | gv_efullname3(tmpstr, gv, Nullch); |
35c1215d |
1532 | DIE(aTHX_ "Undefined sort subroutine \"%"SVf"\" called", |
1533 | tmpstr); |
84d4ea48 |
1534 | } |
1535 | else { |
1536 | DIE(aTHX_ "Undefined subroutine in sort"); |
1537 | } |
1538 | } |
1539 | |
1540 | if (is_xsub) |
1541 | PL_sortcop = (OP*)cv; |
1542 | else { |
1543 | PL_sortcop = CvSTART(cv); |
1544 | SAVEVPTR(CvROOT(cv)->op_ppaddr); |
1545 | CvROOT(cv)->op_ppaddr = PL_ppaddr[OP_NULL]; |
1546 | |
fd617465 |
1547 | SAVECOMPPAD(); |
1548 | PAD_SET_CUR_NOSAVE(CvPADLIST(cv), 1); |
84d4ea48 |
1549 | } |
1550 | } |
1551 | } |
1552 | else { |
1553 | PL_sortcop = Nullop; |
1554 | stash = CopSTASH(PL_curcop); |
1555 | } |
1556 | |
fe1bc4cf |
1557 | /* optimiser converts "@a = sort @a" to "sort \@a"; |
1558 | * in case of tied @a, pessimise: push (@a) onto stack, then assign |
1559 | * result back to @a at the end of this function */ |
0723351e |
1560 | if (priv & OPpSORT_INPLACE) { |
fe1bc4cf |
1561 | assert( MARK+1 == SP && *SP && SvTYPE(*SP) == SVt_PVAV); |
1562 | (void)POPMARK; /* remove mark associated with ex-OP_AASSIGN */ |
1563 | av = (AV*)(*SP); |
1564 | max = AvFILL(av) + 1; |
1565 | if (SvMAGICAL(av)) { |
1566 | MEXTEND(SP, max); |
1567 | p2 = SP; |
fe2774ed |
1568 | for (i=0; i < max; i++) { |
fe1bc4cf |
1569 | SV **svp = av_fetch(av, i, FALSE); |
1570 | *SP++ = (svp) ? *svp : Nullsv; |
1571 | } |
1572 | } |
1573 | else { |
1574 | p1 = p2 = AvARRAY(av); |
1575 | sorting_av = 1; |
1576 | } |
1577 | } |
1578 | else { |
1579 | p2 = MARK+1; |
1580 | max = SP - MARK; |
1581 | } |
1582 | |
0723351e |
1583 | if (priv & OPpSORT_DESCEND) { |
6c3fb703 |
1584 | sortsvp = S_sortsv_desc; |
1585 | } |
1586 | |
83a44efe |
1587 | /* shuffle stack down, removing optional initial cv (p1!=p2), plus |
1588 | * any nulls; also stringify or converting to integer or number as |
1589 | * required any args */ |
fe1bc4cf |
1590 | for (i=max; i > 0 ; i--) { |
1591 | if ((*p1 = *p2++)) { /* Weed out nulls. */ |
1592 | SvTEMP_off(*p1); |
83a44efe |
1593 | if (!PL_sortcop) { |
1594 | if (priv & OPpSORT_NUMERIC) { |
1595 | if (priv & OPpSORT_INTEGER) { |
1596 | if (!SvIOK(*p1)) { |
1597 | if (SvAMAGIC(*p1)) |
1598 | overloading = 1; |
1599 | else |
1600 | (void)sv_2iv(*p1); |
1601 | } |
1602 | } |
1603 | else { |
4d562308 |
1604 | if (!SvNSIOK(*p1)) { |
83a44efe |
1605 | if (SvAMAGIC(*p1)) |
1606 | overloading = 1; |
1607 | else |
1608 | (void)sv_2nv(*p1); |
1609 | } |
4d562308 |
1610 | if (all_SIVs && !SvSIOK(*p1)) |
1611 | all_SIVs = 0; |
83a44efe |
1612 | } |
1613 | } |
1614 | else { |
1615 | if (!SvPOK(*p1)) { |
83a44efe |
1616 | if (SvAMAGIC(*p1)) |
1617 | overloading = 1; |
1618 | else |
83003860 |
1619 | (void)sv_2pv_flags(*p1, 0, |
1620 | SV_GMAGIC|SV_CONST_RETURN); |
83a44efe |
1621 | } |
1622 | } |
84d4ea48 |
1623 | } |
fe1bc4cf |
1624 | p1++; |
84d4ea48 |
1625 | } |
fe1bc4cf |
1626 | else |
1627 | max--; |
84d4ea48 |
1628 | } |
fe1bc4cf |
1629 | if (sorting_av) |
1630 | AvFILLp(av) = max-1; |
1631 | |
1632 | if (max > 1) { |
471178c0 |
1633 | SV **start; |
fe1bc4cf |
1634 | if (PL_sortcop) { |
84d4ea48 |
1635 | PERL_CONTEXT *cx; |
1636 | SV** newsp; |
901017d6 |
1637 | const bool oldcatch = CATCH_GET; |
84d4ea48 |
1638 | |
1639 | SAVETMPS; |
1640 | SAVEOP(); |
1641 | |
1642 | CATCH_SET(TRUE); |
1643 | PUSHSTACKi(PERLSI_SORT); |
1644 | if (!hasargs && !is_xsub) { |
1645 | if (PL_sortstash != stash || !PL_firstgv || !PL_secondgv) { |
1646 | SAVESPTR(PL_firstgv); |
1647 | SAVESPTR(PL_secondgv); |
1648 | PL_firstgv = gv_fetchpv("a", TRUE, SVt_PV); |
1649 | PL_secondgv = gv_fetchpv("b", TRUE, SVt_PV); |
1650 | PL_sortstash = stash; |
1651 | } |
84d4ea48 |
1652 | SAVESPTR(GvSV(PL_firstgv)); |
1653 | SAVESPTR(GvSV(PL_secondgv)); |
1654 | } |
1655 | |
1656 | PUSHBLOCK(cx, CXt_NULL, PL_stack_base); |
471178c0 |
1657 | if (!(flags & OPf_SPECIAL)) { |
84d4ea48 |
1658 | cx->cx_type = CXt_SUB; |
1659 | cx->blk_gimme = G_SCALAR; |
1660 | PUSHSUB(cx); |
84d4ea48 |
1661 | } |
1662 | PL_sortcxix = cxstack_ix; |
1663 | |
1664 | if (hasargs && !is_xsub) { |
1665 | /* This is mostly copied from pp_entersub */ |
dd2155a4 |
1666 | AV *av = (AV*)PAD_SVl(0); |
84d4ea48 |
1667 | |
84d4ea48 |
1668 | cx->blk_sub.savearray = GvAV(PL_defgv); |
1669 | GvAV(PL_defgv) = (AV*)SvREFCNT_inc(av); |
dd2155a4 |
1670 | CX_CURPAD_SAVE(cx->blk_sub); |
84d4ea48 |
1671 | cx->blk_sub.argarray = av; |
1672 | } |
471178c0 |
1673 | |
1674 | start = p1 - max; |
1675 | sortsvp(aTHX_ start, max, |
1676 | is_xsub ? sortcv_xsub : hasargs ? sortcv_stacked : sortcv); |
84d4ea48 |
1677 | |
1678 | POPBLOCK(cx,PL_curpm); |
1679 | PL_stack_sp = newsp; |
1680 | POPSTACK; |
1681 | CATCH_SET(oldcatch); |
1682 | } |
fe1bc4cf |
1683 | else { |
84d4ea48 |
1684 | MEXTEND(SP, 20); /* Can't afford stack realloc on signal. */ |
471178c0 |
1685 | start = sorting_av ? AvARRAY(av) : ORIGMARK+1; |
1686 | sortsvp(aTHX_ start, max, |
0723351e |
1687 | (priv & OPpSORT_NUMERIC) |
4d562308 |
1688 | ? ( ( ( priv & OPpSORT_INTEGER) || all_SIVs) |
84d4ea48 |
1689 | ? ( overloading ? amagic_i_ncmp : sv_i_ncmp) |
4d562308 |
1690 | : ( overloading ? amagic_ncmp : sv_ncmp ) ) |
84d4ea48 |
1691 | : ( IN_LOCALE_RUNTIME |
1692 | ? ( overloading |
1693 | ? amagic_cmp_locale |
1694 | : sv_cmp_locale_static) |
1695 | : ( overloading ? amagic_cmp : sv_cmp_static))); |
471178c0 |
1696 | } |
0723351e |
1697 | if (priv & OPpSORT_REVERSE) { |
471178c0 |
1698 | SV **q = start+max-1; |
1699 | while (start < q) { |
1700 | SV *tmp = *start; |
1701 | *start++ = *q; |
1702 | *q-- = tmp; |
84d4ea48 |
1703 | } |
1704 | } |
1705 | } |
fe1bc4cf |
1706 | if (av && !sorting_av) { |
1707 | /* simulate pp_aassign of tied AV */ |
901017d6 |
1708 | SV** const base = ORIGMARK+1; |
1709 | for (i=0; i < max; i++) { |
1710 | base[i] = newSVsv(base[i]); |
fe1bc4cf |
1711 | } |
1712 | av_clear(av); |
1713 | av_extend(av, max); |
1714 | for (i=0; i < max; i++) { |
901017d6 |
1715 | SV * const sv = base[i]; |
1716 | SV **didstore = av_store(av, i, sv); |
fe1bc4cf |
1717 | if (SvSMAGICAL(sv)) |
1718 | mg_set(sv); |
1719 | if (!didstore) |
1720 | sv_2mortal(sv); |
1721 | } |
1722 | } |
84d4ea48 |
1723 | LEAVE; |
fe1bc4cf |
1724 | PL_stack_sp = ORIGMARK + (sorting_av ? 0 : max); |
84d4ea48 |
1725 | return nextop; |
1726 | } |
1727 | |
1728 | static I32 |
1729 | sortcv(pTHX_ SV *a, SV *b) |
1730 | { |
27da23d5 |
1731 | dVAR; |
901017d6 |
1732 | const I32 oldsaveix = PL_savestack_ix; |
1733 | const I32 oldscopeix = PL_scopestack_ix; |
84d4ea48 |
1734 | I32 result; |
1735 | GvSV(PL_firstgv) = a; |
1736 | GvSV(PL_secondgv) = b; |
1737 | PL_stack_sp = PL_stack_base; |
1738 | PL_op = PL_sortcop; |
1739 | CALLRUNOPS(aTHX); |
1740 | if (PL_stack_sp != PL_stack_base + 1) |
1741 | Perl_croak(aTHX_ "Sort subroutine didn't return single value"); |
1742 | if (!SvNIOKp(*PL_stack_sp)) |
1743 | Perl_croak(aTHX_ "Sort subroutine didn't return a numeric value"); |
1744 | result = SvIV(*PL_stack_sp); |
1745 | while (PL_scopestack_ix > oldscopeix) { |
1746 | LEAVE; |
1747 | } |
1748 | leave_scope(oldsaveix); |
1749 | return result; |
1750 | } |
1751 | |
1752 | static I32 |
1753 | sortcv_stacked(pTHX_ SV *a, SV *b) |
1754 | { |
27da23d5 |
1755 | dVAR; |
901017d6 |
1756 | const I32 oldsaveix = PL_savestack_ix; |
1757 | const I32 oldscopeix = PL_scopestack_ix; |
84d4ea48 |
1758 | I32 result; |
901017d6 |
1759 | AV * const av = GvAV(PL_defgv); |
84d4ea48 |
1760 | |
1761 | if (AvMAX(av) < 1) { |
1762 | SV** ary = AvALLOC(av); |
1763 | if (AvARRAY(av) != ary) { |
1764 | AvMAX(av) += AvARRAY(av) - AvALLOC(av); |
f880fe2f |
1765 | SvPV_set(av, (char*)ary); |
84d4ea48 |
1766 | } |
1767 | if (AvMAX(av) < 1) { |
1768 | AvMAX(av) = 1; |
1769 | Renew(ary,2,SV*); |
f880fe2f |
1770 | SvPV_set(av, (char*)ary); |
84d4ea48 |
1771 | } |
1772 | } |
1773 | AvFILLp(av) = 1; |
1774 | |
1775 | AvARRAY(av)[0] = a; |
1776 | AvARRAY(av)[1] = b; |
1777 | PL_stack_sp = PL_stack_base; |
1778 | PL_op = PL_sortcop; |
1779 | CALLRUNOPS(aTHX); |
1780 | if (PL_stack_sp != PL_stack_base + 1) |
1781 | Perl_croak(aTHX_ "Sort subroutine didn't return single value"); |
1782 | if (!SvNIOKp(*PL_stack_sp)) |
1783 | Perl_croak(aTHX_ "Sort subroutine didn't return a numeric value"); |
1784 | result = SvIV(*PL_stack_sp); |
1785 | while (PL_scopestack_ix > oldscopeix) { |
1786 | LEAVE; |
1787 | } |
1788 | leave_scope(oldsaveix); |
1789 | return result; |
1790 | } |
1791 | |
1792 | static I32 |
1793 | sortcv_xsub(pTHX_ SV *a, SV *b) |
1794 | { |
27da23d5 |
1795 | dVAR; dSP; |
901017d6 |
1796 | const I32 oldsaveix = PL_savestack_ix; |
1797 | const I32 oldscopeix = PL_scopestack_ix; |
1798 | CV * const cv=(CV*)PL_sortcop; |
84d4ea48 |
1799 | I32 result; |
84d4ea48 |
1800 | |
1801 | SP = PL_stack_base; |
1802 | PUSHMARK(SP); |
1803 | EXTEND(SP, 2); |
1804 | *++SP = a; |
1805 | *++SP = b; |
1806 | PUTBACK; |
1807 | (void)(*CvXSUB(cv))(aTHX_ cv); |
1808 | if (PL_stack_sp != PL_stack_base + 1) |
1809 | Perl_croak(aTHX_ "Sort subroutine didn't return single value"); |
1810 | if (!SvNIOKp(*PL_stack_sp)) |
1811 | Perl_croak(aTHX_ "Sort subroutine didn't return a numeric value"); |
1812 | result = SvIV(*PL_stack_sp); |
1813 | while (PL_scopestack_ix > oldscopeix) { |
1814 | LEAVE; |
1815 | } |
1816 | leave_scope(oldsaveix); |
1817 | return result; |
1818 | } |
1819 | |
1820 | |
1821 | static I32 |
1822 | sv_ncmp(pTHX_ SV *a, SV *b) |
1823 | { |
901017d6 |
1824 | const NV nv1 = SvNSIV(a); |
1825 | const NV nv2 = SvNSIV(b); |
84d4ea48 |
1826 | return nv1 < nv2 ? -1 : nv1 > nv2 ? 1 : 0; |
1827 | } |
1828 | |
1829 | static I32 |
1830 | sv_i_ncmp(pTHX_ SV *a, SV *b) |
1831 | { |
901017d6 |
1832 | const IV iv1 = SvIV(a); |
1833 | const IV iv2 = SvIV(b); |
84d4ea48 |
1834 | return iv1 < iv2 ? -1 : iv1 > iv2 ? 1 : 0; |
1835 | } |
901017d6 |
1836 | |
1837 | #define tryCALL_AMAGICbin(left,right,meth) \ |
1838 | (PL_amagic_generation && (SvAMAGIC(left)||SvAMAGIC(right))) \ |
1839 | ? amagic_call(left, right, CAT2(meth,_amg), 0) \ |
1840 | : Nullsv; |
84d4ea48 |
1841 | |
1842 | static I32 |
1843 | amagic_ncmp(pTHX_ register SV *a, register SV *b) |
1844 | { |
901017d6 |
1845 | SV * const tmpsv = tryCALL_AMAGICbin(a,b,ncmp); |
84d4ea48 |
1846 | if (tmpsv) { |
84d4ea48 |
1847 | if (SvIOK(tmpsv)) { |
901017d6 |
1848 | const I32 i = SvIVX(tmpsv); |
84d4ea48 |
1849 | if (i > 0) |
1850 | return 1; |
1851 | return i? -1 : 0; |
1852 | } |
901017d6 |
1853 | else { |
1854 | const NV d = SvNV(tmpsv); |
1855 | if (d > 0) |
1856 | return 1; |
1857 | return d ? -1 : 0; |
1858 | } |
84d4ea48 |
1859 | } |
1860 | return sv_ncmp(aTHX_ a, b); |
1861 | } |
1862 | |
1863 | static I32 |
1864 | amagic_i_ncmp(pTHX_ register SV *a, register SV *b) |
1865 | { |
901017d6 |
1866 | SV * const tmpsv = tryCALL_AMAGICbin(a,b,ncmp); |
84d4ea48 |
1867 | if (tmpsv) { |
84d4ea48 |
1868 | if (SvIOK(tmpsv)) { |
901017d6 |
1869 | const I32 i = SvIVX(tmpsv); |
84d4ea48 |
1870 | if (i > 0) |
1871 | return 1; |
1872 | return i? -1 : 0; |
1873 | } |
901017d6 |
1874 | else { |
1875 | const NV d = SvNV(tmpsv); |
1876 | if (d > 0) |
1877 | return 1; |
1878 | return d ? -1 : 0; |
1879 | } |
84d4ea48 |
1880 | } |
1881 | return sv_i_ncmp(aTHX_ a, b); |
1882 | } |
1883 | |
1884 | static I32 |
1885 | amagic_cmp(pTHX_ register SV *str1, register SV *str2) |
1886 | { |
901017d6 |
1887 | SV * const tmpsv = tryCALL_AMAGICbin(str1,str2,scmp); |
84d4ea48 |
1888 | if (tmpsv) { |
84d4ea48 |
1889 | if (SvIOK(tmpsv)) { |
901017d6 |
1890 | const I32 i = SvIVX(tmpsv); |
84d4ea48 |
1891 | if (i > 0) |
1892 | return 1; |
1893 | return i? -1 : 0; |
1894 | } |
901017d6 |
1895 | else { |
1896 | const NV d = SvNV(tmpsv); |
1897 | if (d > 0) |
1898 | return 1; |
1899 | return d? -1 : 0; |
1900 | } |
84d4ea48 |
1901 | } |
1902 | return sv_cmp(str1, str2); |
1903 | } |
1904 | |
1905 | static I32 |
1906 | amagic_cmp_locale(pTHX_ register SV *str1, register SV *str2) |
1907 | { |
901017d6 |
1908 | SV * const tmpsv = tryCALL_AMAGICbin(str1,str2,scmp); |
84d4ea48 |
1909 | if (tmpsv) { |
84d4ea48 |
1910 | if (SvIOK(tmpsv)) { |
901017d6 |
1911 | const I32 i = SvIVX(tmpsv); |
84d4ea48 |
1912 | if (i > 0) |
1913 | return 1; |
1914 | return i? -1 : 0; |
1915 | } |
901017d6 |
1916 | else { |
1917 | const NV d = SvNV(tmpsv); |
1918 | if (d > 0) |
1919 | return 1; |
1920 | return d? -1 : 0; |
1921 | } |
84d4ea48 |
1922 | } |
1923 | return sv_cmp_locale(str1, str2); |
1924 | } |
241d1a3b |
1925 | |
1926 | /* |
1927 | * Local variables: |
1928 | * c-indentation-style: bsd |
1929 | * c-basic-offset: 4 |
1930 | * indent-tabs-mode: t |
1931 | * End: |
1932 | * |
37442d52 |
1933 | * ex: set ts=8 sts=4 sw=4 noet: |
1934 | */ |