Commit | Line | Data |
---|---|---|
84d4ea48 JH |
1 | /* pp_sort.c |
2 | * | |
1129b882 NC |
3 | * Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, |
4 | * 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008 by Larry Wall and others | |
84d4ea48 JH |
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 | /* | |
4ac71550 TC |
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 | * [p.931 of _The Lord of the Rings_, VI/ii: "The Land of Shadow"] | |
84d4ea48 JH |
16 | */ |
17 | ||
166f8a29 DM |
18 | /* This file contains pp ("push/pop") functions that |
19 | * execute the opcodes that make up a perl program. A typical pp function | |
20 | * expects to find its arguments on the stack, and usually pushes its | |
21 | * results onto the stack, hence the 'pp' terminology. Each OP structure | |
22 | * contains a pointer to the relevant pp_foo() function. | |
23 | * | |
24 | * This particular file just contains pp_sort(), which is complex | |
25 | * enough to merit its own file! See the other pp*.c files for the rest of | |
26 | * the pp_ functions. | |
27 | */ | |
28 | ||
84d4ea48 JH |
29 | #include "EXTERN.h" |
30 | #define PERL_IN_PP_SORT_C | |
31 | #include "perl.h" | |
32 | ||
84d4ea48 JH |
33 | #define sv_cmp_static Perl_sv_cmp |
34 | #define sv_cmp_locale_static Perl_sv_cmp_locale | |
35 | ||
c53fc8a6 JH |
36 | #ifndef SMALLSORT |
37 | #define SMALLSORT (200) | |
38 | #endif | |
39 | ||
7b9ef140 RH |
40 | /* Flags for qsortsv and mergesortsv */ |
41 | #define SORTf_DESC 1 | |
42 | #define SORTf_STABLE 2 | |
afe59f35 | 43 | #define SORTf_UNSTABLE 8 |
7b9ef140 | 44 | |
84d4ea48 JH |
45 | /* |
46 | * The mergesort implementation is by Peter M. Mcilroy <pmcilroy@lucent.com>. | |
47 | * | |
48 | * The original code was written in conjunction with BSD Computer Software | |
49 | * Research Group at University of California, Berkeley. | |
50 | * | |
393db44d JL |
51 | * See also: "Optimistic Sorting and Information Theoretic Complexity" |
52 | * Peter McIlroy | |
53 | * SODA (Fourth Annual ACM-SIAM Symposium on Discrete Algorithms), | |
54 | * pp 467-474, Austin, Texas, 25-27 January 1993. | |
84d4ea48 | 55 | * |
393db44d | 56 | * The integration to Perl is by John P. Linderman <jpl.jpl@gmail.com>. |
84d4ea48 JH |
57 | * |
58 | * The code can be distributed under the same terms as Perl itself. | |
59 | * | |
60 | */ | |
61 | ||
84d4ea48 JH |
62 | |
63 | typedef char * aptr; /* pointer for arithmetic on sizes */ | |
64 | typedef SV * gptr; /* pointers in our lists */ | |
65 | ||
66 | /* Binary merge internal sort, with a few special mods | |
67 | ** for the special perl environment it now finds itself in. | |
68 | ** | |
69 | ** Things that were once options have been hotwired | |
70 | ** to values suitable for this use. In particular, we'll always | |
71 | ** initialize looking for natural runs, we'll always produce stable | |
72 | ** output, and we'll always do Peter McIlroy's binary merge. | |
73 | */ | |
74 | ||
75 | /* Pointer types for arithmetic and storage and convenience casts */ | |
76 | ||
77 | #define APTR(P) ((aptr)(P)) | |
78 | #define GPTP(P) ((gptr *)(P)) | |
79 | #define GPPP(P) ((gptr **)(P)) | |
80 | ||
81 | ||
82 | /* byte offset from pointer P to (larger) pointer Q */ | |
83 | #define BYTEOFF(P, Q) (APTR(Q) - APTR(P)) | |
84 | ||
85 | #define PSIZE sizeof(gptr) | |
86 | ||
87 | /* If PSIZE is power of 2, make PSHIFT that power, if that helps */ | |
88 | ||
89 | #ifdef PSHIFT | |
90 | #define PNELEM(P, Q) (BYTEOFF(P,Q) >> (PSHIFT)) | |
91 | #define PNBYTE(N) ((N) << (PSHIFT)) | |
92 | #define PINDEX(P, N) (GPTP(APTR(P) + PNBYTE(N))) | |
93 | #else | |
94 | /* Leave optimization to compiler */ | |
95 | #define PNELEM(P, Q) (GPTP(Q) - GPTP(P)) | |
96 | #define PNBYTE(N) ((N) * (PSIZE)) | |
97 | #define PINDEX(P, N) (GPTP(P) + (N)) | |
98 | #endif | |
99 | ||
100 | /* Pointer into other corresponding to pointer into this */ | |
101 | #define POTHER(P, THIS, OTHER) GPTP(APTR(OTHER) + BYTEOFF(THIS,P)) | |
102 | ||
103 | #define FROMTOUPTO(src, dst, lim) do *dst++ = *src++; while(src<lim) | |
104 | ||
105 | ||
486ec47a | 106 | /* Runs are identified by a pointer in the auxiliary list. |
84d4ea48 JH |
107 | ** The pointer is at the start of the list, |
108 | ** and it points to the start of the next list. | |
109 | ** NEXT is used as an lvalue, too. | |
110 | */ | |
111 | ||
112 | #define NEXT(P) (*GPPP(P)) | |
113 | ||
114 | ||
115 | /* PTHRESH is the minimum number of pairs with the same sense to justify | |
116 | ** checking for a run and extending it. Note that PTHRESH counts PAIRS, | |
117 | ** not just elements, so PTHRESH == 8 means a run of 16. | |
118 | */ | |
119 | ||
120 | #define PTHRESH (8) | |
121 | ||
122 | /* RTHRESH is the number of elements in a run that must compare low | |
123 | ** to the low element from the opposing run before we justify | |
124 | ** doing a binary rampup instead of single stepping. | |
125 | ** In random input, N in a row low should only happen with | |
126 | ** probability 2^(1-N), so we can risk that we are dealing | |
127 | ** with orderly input without paying much when we aren't. | |
128 | */ | |
129 | ||
130 | #define RTHRESH (6) | |
131 | ||
132 | ||
133 | /* | |
134 | ** Overview of algorithm and variables. | |
135 | ** The array of elements at list1 will be organized into runs of length 2, | |
136 | ** or runs of length >= 2 * PTHRESH. We only try to form long runs when | |
137 | ** PTHRESH adjacent pairs compare in the same way, suggesting overall order. | |
138 | ** | |
139 | ** Unless otherwise specified, pair pointers address the first of two elements. | |
140 | ** | |
a0288114 AL |
141 | ** b and b+1 are a pair that compare with sense "sense". |
142 | ** b is the "bottom" of adjacent pairs that might form a longer run. | |
84d4ea48 JH |
143 | ** |
144 | ** p2 parallels b in the list2 array, where runs are defined by | |
145 | ** a pointer chain. | |
146 | ** | |
a0288114 | 147 | ** t represents the "top" of the adjacent pairs that might extend |
84d4ea48 JH |
148 | ** the run beginning at b. Usually, t addresses a pair |
149 | ** that compares with opposite sense from (b,b+1). | |
150 | ** However, it may also address a singleton element at the end of list1, | |
a0288114 | 151 | ** or it may be equal to "last", the first element beyond list1. |
84d4ea48 JH |
152 | ** |
153 | ** r addresses the Nth pair following b. If this would be beyond t, | |
154 | ** we back it off to t. Only when r is less than t do we consider the | |
155 | ** run long enough to consider checking. | |
156 | ** | |
157 | ** q addresses a pair such that the pairs at b through q already form a run. | |
158 | ** Often, q will equal b, indicating we only are sure of the pair itself. | |
159 | ** However, a search on the previous cycle may have revealed a longer run, | |
160 | ** so q may be greater than b. | |
161 | ** | |
162 | ** p is used to work back from a candidate r, trying to reach q, | |
163 | ** which would mean b through r would be a run. If we discover such a run, | |
164 | ** we start q at r and try to push it further towards t. | |
165 | ** If b through r is NOT a run, we detect the wrong order at (p-1,p). | |
166 | ** In any event, after the check (if any), we have two main cases. | |
167 | ** | |
168 | ** 1) Short run. b <= q < p <= r <= t. | |
169 | ** b through q is a run (perhaps trivial) | |
170 | ** q through p are uninteresting pairs | |
171 | ** p through r is a run | |
172 | ** | |
173 | ** 2) Long run. b < r <= q < t. | |
174 | ** b through q is a run (of length >= 2 * PTHRESH) | |
175 | ** | |
176 | ** Note that degenerate cases are not only possible, but likely. | |
177 | ** For example, if the pair following b compares with opposite sense, | |
178 | ** then b == q < p == r == t. | |
179 | */ | |
180 | ||
181 | ||
957d8989 | 182 | static IV |
d4c19fe8 | 183 | dynprep(pTHX_ gptr *list1, gptr *list2, size_t nmemb, const SVCOMPARE_t cmp) |
84d4ea48 | 184 | { |
957d8989 | 185 | I32 sense; |
eb578fdb KW |
186 | gptr *b, *p, *q, *t, *p2; |
187 | gptr *last, *r; | |
957d8989 | 188 | IV runs = 0; |
84d4ea48 JH |
189 | |
190 | b = list1; | |
191 | last = PINDEX(b, nmemb); | |
192 | sense = (cmp(aTHX_ *b, *(b+1)) > 0); | |
193 | for (p2 = list2; b < last; ) { | |
194 | /* We just started, or just reversed sense. | |
195 | ** Set t at end of pairs with the prevailing sense. | |
196 | */ | |
197 | for (p = b+2, t = p; ++p < last; t = ++p) { | |
198 | if ((cmp(aTHX_ *t, *p) > 0) != sense) break; | |
199 | } | |
200 | q = b; | |
201 | /* Having laid out the playing field, look for long runs */ | |
202 | do { | |
203 | p = r = b + (2 * PTHRESH); | |
204 | if (r >= t) p = r = t; /* too short to care about */ | |
205 | else { | |
206 | while (((cmp(aTHX_ *(p-1), *p) > 0) == sense) && | |
47127b64 | 207 | ((p -= 2) > q)) {} |
84d4ea48 JH |
208 | if (p <= q) { |
209 | /* b through r is a (long) run. | |
210 | ** Extend it as far as possible. | |
211 | */ | |
212 | p = q = r; | |
213 | while (((p += 2) < t) && | |
214 | ((cmp(aTHX_ *(p-1), *p) > 0) == sense)) q = p; | |
215 | r = p = q + 2; /* no simple pairs, no after-run */ | |
216 | } | |
217 | } | |
218 | if (q > b) { /* run of greater than 2 at b */ | |
d4c19fe8 AL |
219 | gptr *savep = p; |
220 | ||
84d4ea48 JH |
221 | p = q += 2; |
222 | /* pick up singleton, if possible */ | |
223 | if ((p == t) && | |
224 | ((t + 1) == last) && | |
225 | ((cmp(aTHX_ *(p-1), *p) > 0) == sense)) | |
226 | savep = r = p = q = last; | |
957d8989 | 227 | p2 = NEXT(p2) = p2 + (p - b); ++runs; |
d4c19fe8 AL |
228 | if (sense) |
229 | while (b < --p) { | |
230 | const gptr c = *b; | |
231 | *b++ = *p; | |
232 | *p = c; | |
233 | } | |
84d4ea48 JH |
234 | p = savep; |
235 | } | |
236 | while (q < p) { /* simple pairs */ | |
957d8989 | 237 | p2 = NEXT(p2) = p2 + 2; ++runs; |
84d4ea48 | 238 | if (sense) { |
d4c19fe8 | 239 | const gptr c = *q++; |
84d4ea48 JH |
240 | *(q-1) = *q; |
241 | *q++ = c; | |
242 | } else q += 2; | |
243 | } | |
244 | if (((b = p) == t) && ((t+1) == last)) { | |
957d8989 | 245 | NEXT(p2) = p2 + 1; ++runs; |
84d4ea48 JH |
246 | b++; |
247 | } | |
248 | q = r; | |
249 | } while (b < t); | |
250 | sense = !sense; | |
251 | } | |
957d8989 | 252 | return runs; |
84d4ea48 JH |
253 | } |
254 | ||
255 | ||
3fe0b9a9 | 256 | /* The original merge sort, in use since 5.7, was as fast as, or faster than, |
957d8989 | 257 | * qsort on many platforms, but slower than qsort, conspicuously so, |
3fe0b9a9 | 258 | * on others. The most likely explanation was platform-specific |
957d8989 JL |
259 | * differences in cache sizes and relative speeds. |
260 | * | |
261 | * The quicksort divide-and-conquer algorithm guarantees that, as the | |
262 | * problem is subdivided into smaller and smaller parts, the parts | |
263 | * fit into smaller (and faster) caches. So it doesn't matter how | |
264 | * many levels of cache exist, quicksort will "find" them, and, | |
e62b3022 | 265 | * as long as smaller is faster, take advantage of them. |
957d8989 | 266 | * |
3fe0b9a9 | 267 | * By contrast, consider how the original mergesort algorithm worked. |
957d8989 JL |
268 | * Suppose we have five runs (each typically of length 2 after dynprep). |
269 | * | |
270 | * pass base aux | |
271 | * 0 1 2 3 4 5 | |
272 | * 1 12 34 5 | |
273 | * 2 1234 5 | |
274 | * 3 12345 | |
275 | * 4 12345 | |
276 | * | |
277 | * Adjacent pairs are merged in "grand sweeps" through the input. | |
278 | * This means, on pass 1, the records in runs 1 and 2 aren't revisited until | |
279 | * runs 3 and 4 are merged and the runs from run 5 have been copied. | |
280 | * The only cache that matters is one large enough to hold *all* the input. | |
281 | * On some platforms, this may be many times slower than smaller caches. | |
282 | * | |
283 | * The following pseudo-code uses the same basic merge algorithm, | |
284 | * but in a divide-and-conquer way. | |
285 | * | |
286 | * # merge $runs runs at offset $offset of list $list1 into $list2. | |
287 | * # all unmerged runs ($runs == 1) originate in list $base. | |
288 | * sub mgsort2 { | |
289 | * my ($offset, $runs, $base, $list1, $list2) = @_; | |
290 | * | |
291 | * if ($runs == 1) { | |
292 | * if ($list1 is $base) copy run to $list2 | |
293 | * return offset of end of list (or copy) | |
294 | * } else { | |
295 | * $off2 = mgsort2($offset, $runs-($runs/2), $base, $list2, $list1) | |
296 | * mgsort2($off2, $runs/2, $base, $list2, $list1) | |
297 | * merge the adjacent runs at $offset of $list1 into $list2 | |
298 | * return the offset of the end of the merged runs | |
299 | * } | |
300 | * } | |
301 | * mgsort2(0, $runs, $base, $aux, $base); | |
302 | * | |
303 | * For our 5 runs, the tree of calls looks like | |
304 | * | |
305 | * 5 | |
306 | * 3 2 | |
307 | * 2 1 1 1 | |
308 | * 1 1 | |
309 | * | |
310 | * 1 2 3 4 5 | |
311 | * | |
312 | * and the corresponding activity looks like | |
313 | * | |
314 | * copy runs 1 and 2 from base to aux | |
315 | * merge runs 1 and 2 from aux to base | |
316 | * (run 3 is where it belongs, no copy needed) | |
317 | * merge runs 12 and 3 from base to aux | |
318 | * (runs 4 and 5 are where they belong, no copy needed) | |
319 | * merge runs 4 and 5 from base to aux | |
320 | * merge runs 123 and 45 from aux to base | |
321 | * | |
322 | * Note that we merge runs 1 and 2 immediately after copying them, | |
323 | * while they are still likely to be in fast cache. Similarly, | |
324 | * run 3 is merged with run 12 while it still may be lingering in cache. | |
325 | * This implementation should therefore enjoy much of the cache-friendly | |
326 | * behavior that quicksort does. In addition, it does less copying | |
327 | * than the original mergesort implementation (only runs 1 and 2 are copied) | |
328 | * and the "balancing" of merges is better (merged runs comprise more nearly | |
329 | * equal numbers of original runs). | |
330 | * | |
331 | * The actual cache-friendly implementation will use a pseudo-stack | |
332 | * to avoid recursion, and will unroll processing of runs of length 2, | |
333 | * but it is otherwise similar to the recursive implementation. | |
957d8989 JL |
334 | */ |
335 | ||
336 | typedef struct { | |
337 | IV offset; /* offset of 1st of 2 runs at this level */ | |
338 | IV runs; /* how many runs must be combined into 1 */ | |
339 | } off_runs; /* pseudo-stack element */ | |
340 | ||
6c3fb703 NC |
341 | |
342 | static I32 | |
31e9e0a3 | 343 | cmp_desc(pTHX_ gptr const a, gptr const b) |
6c3fb703 NC |
344 | { |
345 | return -PL_sort_RealCmp(aTHX_ a, b); | |
346 | } | |
347 | ||
e2091bb6 | 348 | /* |
83f1d9e7 KW |
349 | =head1 SV Manipulation Functions |
350 | ||
e2091bb6 Z |
351 | =for apidoc sortsv_flags |
352 | ||
353 | In-place sort an array of SV pointers with the given comparison routine, | |
354 | with various SORTf_* flag options. | |
355 | ||
356 | =cut | |
357 | */ | |
358 | void | |
359 | Perl_sortsv_flags(pTHX_ gptr *base, size_t nmemb, SVCOMPARE_t cmp, U32 flags) | |
957d8989 | 360 | { |
551405c4 | 361 | IV i, run, offset; |
957d8989 | 362 | I32 sense, level; |
eb578fdb | 363 | gptr *f1, *f2, *t, *b, *p; |
957d8989 | 364 | int iwhich; |
551405c4 | 365 | gptr *aux; |
957d8989 JL |
366 | gptr *p1; |
367 | gptr small[SMALLSORT]; | |
368 | gptr *which[3]; | |
369 | off_runs stack[60], *stackp; | |
d4c19fe8 | 370 | SVCOMPARE_t savecmp = NULL; |
957d8989 | 371 | |
e2091bb6 | 372 | PERL_ARGS_ASSERT_SORTSV_FLAGS; |
957d8989 | 373 | if (nmemb <= 1) return; /* sorted trivially */ |
6c3fb703 | 374 | |
f4f44d65 | 375 | if ((flags & SORTf_DESC) != 0) { |
6c3fb703 NC |
376 | savecmp = PL_sort_RealCmp; /* Save current comparison routine, if any */ |
377 | PL_sort_RealCmp = cmp; /* Put comparison routine where cmp_desc can find it */ | |
378 | cmp = cmp_desc; | |
379 | } | |
380 | ||
957d8989 | 381 | if (nmemb <= SMALLSORT) aux = small; /* use stack for aux array */ |
486ec47a | 382 | else { Newx(aux,nmemb,gptr); } /* allocate auxiliary array */ |
957d8989 JL |
383 | level = 0; |
384 | stackp = stack; | |
385 | stackp->runs = dynprep(aTHX_ base, aux, nmemb, cmp); | |
386 | stackp->offset = offset = 0; | |
387 | which[0] = which[2] = base; | |
388 | which[1] = aux; | |
389 | for (;;) { | |
390 | /* On levels where both runs have be constructed (stackp->runs == 0), | |
391 | * merge them, and note the offset of their end, in case the offset | |
392 | * is needed at the next level up. Hop up a level, and, | |
393 | * as long as stackp->runs is 0, keep merging. | |
394 | */ | |
551405c4 AL |
395 | IV runs = stackp->runs; |
396 | if (runs == 0) { | |
397 | gptr *list1, *list2; | |
957d8989 JL |
398 | iwhich = level & 1; |
399 | list1 = which[iwhich]; /* area where runs are now */ | |
400 | list2 = which[++iwhich]; /* area for merged runs */ | |
401 | do { | |
eb578fdb | 402 | gptr *l1, *l2, *tp2; |
957d8989 JL |
403 | offset = stackp->offset; |
404 | f1 = p1 = list1 + offset; /* start of first run */ | |
405 | p = tp2 = list2 + offset; /* where merged run will go */ | |
406 | t = NEXT(p); /* where first run ends */ | |
407 | f2 = l1 = POTHER(t, list2, list1); /* ... on the other side */ | |
408 | t = NEXT(t); /* where second runs ends */ | |
409 | l2 = POTHER(t, list2, list1); /* ... on the other side */ | |
410 | offset = PNELEM(list2, t); | |
411 | while (f1 < l1 && f2 < l2) { | |
412 | /* If head 1 is larger than head 2, find ALL the elements | |
413 | ** in list 2 strictly less than head1, write them all, | |
414 | ** then head 1. Then compare the new heads, and repeat, | |
415 | ** until one or both lists are exhausted. | |
416 | ** | |
417 | ** In all comparisons (after establishing | |
418 | ** which head to merge) the item to merge | |
419 | ** (at pointer q) is the first operand of | |
420 | ** the comparison. When we want to know | |
a0288114 | 421 | ** if "q is strictly less than the other", |
957d8989 JL |
422 | ** we can't just do |
423 | ** cmp(q, other) < 0 | |
424 | ** because stability demands that we treat equality | |
425 | ** as high when q comes from l2, and as low when | |
426 | ** q was from l1. So we ask the question by doing | |
427 | ** cmp(q, other) <= sense | |
428 | ** and make sense == 0 when equality should look low, | |
429 | ** and -1 when equality should look high. | |
430 | */ | |
431 | ||
eb578fdb | 432 | gptr *q; |
957d8989 JL |
433 | if (cmp(aTHX_ *f1, *f2) <= 0) { |
434 | q = f2; b = f1; t = l1; | |
435 | sense = -1; | |
436 | } else { | |
437 | q = f1; b = f2; t = l2; | |
438 | sense = 0; | |
439 | } | |
440 | ||
441 | ||
442 | /* ramp up | |
443 | ** | |
444 | ** Leave t at something strictly | |
445 | ** greater than q (or at the end of the list), | |
446 | ** and b at something strictly less than q. | |
447 | */ | |
448 | for (i = 1, run = 0 ;;) { | |
449 | if ((p = PINDEX(b, i)) >= t) { | |
450 | /* off the end */ | |
451 | if (((p = PINDEX(t, -1)) > b) && | |
452 | (cmp(aTHX_ *q, *p) <= sense)) | |
453 | t = p; | |
454 | else b = p; | |
455 | break; | |
456 | } else if (cmp(aTHX_ *q, *p) <= sense) { | |
457 | t = p; | |
458 | break; | |
459 | } else b = p; | |
460 | if (++run >= RTHRESH) i += i; | |
461 | } | |
462 | ||
463 | ||
464 | /* q is known to follow b and must be inserted before t. | |
465 | ** Increment b, so the range of possibilities is [b,t). | |
466 | ** Round binary split down, to favor early appearance. | |
467 | ** Adjust b and t until q belongs just before t. | |
468 | */ | |
469 | ||
470 | b++; | |
471 | while (b < t) { | |
472 | p = PINDEX(b, (PNELEM(b, t) - 1) / 2); | |
473 | if (cmp(aTHX_ *q, *p) <= sense) { | |
474 | t = p; | |
475 | } else b = p + 1; | |
476 | } | |
477 | ||
478 | ||
479 | /* Copy all the strictly low elements */ | |
480 | ||
481 | if (q == f1) { | |
482 | FROMTOUPTO(f2, tp2, t); | |
483 | *tp2++ = *f1++; | |
484 | } else { | |
485 | FROMTOUPTO(f1, tp2, t); | |
486 | *tp2++ = *f2++; | |
487 | } | |
488 | } | |
489 | ||
490 | ||
491 | /* Run out remaining list */ | |
492 | if (f1 == l1) { | |
493 | if (f2 < l2) FROMTOUPTO(f2, tp2, l2); | |
494 | } else FROMTOUPTO(f1, tp2, l1); | |
495 | p1 = NEXT(p1) = POTHER(tp2, list2, list1); | |
496 | ||
497 | if (--level == 0) goto done; | |
498 | --stackp; | |
499 | t = list1; list1 = list2; list2 = t; /* swap lists */ | |
500 | } while ((runs = stackp->runs) == 0); | |
501 | } | |
502 | ||
503 | ||
504 | stackp->runs = 0; /* current run will finish level */ | |
505 | /* While there are more than 2 runs remaining, | |
506 | * turn them into exactly 2 runs (at the "other" level), | |
507 | * each made up of approximately half the runs. | |
508 | * Stack the second half for later processing, | |
509 | * and set about producing the first half now. | |
510 | */ | |
511 | while (runs > 2) { | |
512 | ++level; | |
513 | ++stackp; | |
514 | stackp->offset = offset; | |
515 | runs -= stackp->runs = runs / 2; | |
516 | } | |
517 | /* We must construct a single run from 1 or 2 runs. | |
518 | * All the original runs are in which[0] == base. | |
519 | * The run we construct must end up in which[level&1]. | |
520 | */ | |
521 | iwhich = level & 1; | |
522 | if (runs == 1) { | |
523 | /* Constructing a single run from a single run. | |
524 | * If it's where it belongs already, there's nothing to do. | |
525 | * Otherwise, copy it to where it belongs. | |
526 | * A run of 1 is either a singleton at level 0, | |
527 | * or the second half of a split 3. In neither event | |
528 | * is it necessary to set offset. It will be set by the merge | |
529 | * that immediately follows. | |
530 | */ | |
531 | if (iwhich) { /* Belongs in aux, currently in base */ | |
532 | f1 = b = PINDEX(base, offset); /* where list starts */ | |
533 | f2 = PINDEX(aux, offset); /* where list goes */ | |
534 | t = NEXT(f2); /* where list will end */ | |
535 | offset = PNELEM(aux, t); /* offset thereof */ | |
536 | t = PINDEX(base, offset); /* where it currently ends */ | |
537 | FROMTOUPTO(f1, f2, t); /* copy */ | |
538 | NEXT(b) = t; /* set up parallel pointer */ | |
539 | } else if (level == 0) goto done; /* single run at level 0 */ | |
540 | } else { | |
541 | /* Constructing a single run from two runs. | |
542 | * The merge code at the top will do that. | |
543 | * We need only make sure the two runs are in the "other" array, | |
544 | * so they'll end up in the correct array after the merge. | |
545 | */ | |
546 | ++level; | |
547 | ++stackp; | |
548 | stackp->offset = offset; | |
549 | stackp->runs = 0; /* take care of both runs, trigger merge */ | |
550 | if (!iwhich) { /* Merged runs belong in aux, copy 1st */ | |
551 | f1 = b = PINDEX(base, offset); /* where first run starts */ | |
552 | f2 = PINDEX(aux, offset); /* where it will be copied */ | |
553 | t = NEXT(f2); /* where first run will end */ | |
554 | offset = PNELEM(aux, t); /* offset thereof */ | |
555 | p = PINDEX(base, offset); /* end of first run */ | |
556 | t = NEXT(t); /* where second run will end */ | |
557 | t = PINDEX(base, PNELEM(aux, t)); /* where it now ends */ | |
558 | FROMTOUPTO(f1, f2, t); /* copy both runs */ | |
486ec47a | 559 | NEXT(b) = p; /* paralleled pointer for 1st */ |
957d8989 JL |
560 | NEXT(p) = t; /* ... and for second */ |
561 | } | |
562 | } | |
563 | } | |
7b52d656 | 564 | done: |
957d8989 | 565 | if (aux != small) Safefree(aux); /* free iff allocated */ |
0e1d050c | 566 | if (savecmp != NULL) { |
6c3fb703 NC |
567 | PL_sort_RealCmp = savecmp; /* Restore current comparison routine, if any */ |
568 | } | |
957d8989 JL |
569 | return; |
570 | } | |
571 | ||
84d4ea48 JH |
572 | /* |
573 | * The quicksort implementation was derived from source code contributed | |
574 | * by Tom Horsley. | |
575 | * | |
576 | * NOTE: this code was derived from Tom Horsley's qsort replacement | |
577 | * and should not be confused with the original code. | |
578 | */ | |
579 | ||
580 | /* Copyright (C) Tom Horsley, 1997. All rights reserved. | |
581 | ||
582 | Permission granted to distribute under the same terms as perl which are | |
583 | (briefly): | |
584 | ||
585 | This program is free software; you can redistribute it and/or modify | |
586 | it under the terms of either: | |
587 | ||
588 | a) the GNU General Public License as published by the Free | |
589 | Software Foundation; either version 1, or (at your option) any | |
590 | later version, or | |
591 | ||
592 | b) the "Artistic License" which comes with this Kit. | |
593 | ||
594 | Details on the perl license can be found in the perl source code which | |
595 | may be located via the www.perl.com web page. | |
596 | ||
597 | This is the most wonderfulest possible qsort I can come up with (and | |
598 | still be mostly portable) My (limited) tests indicate it consistently | |
599 | does about 20% fewer calls to compare than does the qsort in the Visual | |
600 | C++ library, other vendors may vary. | |
601 | ||
602 | Some of the ideas in here can be found in "Algorithms" by Sedgewick, | |
603 | others I invented myself (or more likely re-invented since they seemed | |
604 | pretty obvious once I watched the algorithm operate for a while). | |
605 | ||
606 | Most of this code was written while watching the Marlins sweep the Giants | |
607 | in the 1997 National League Playoffs - no Braves fans allowed to use this | |
608 | code (just kidding :-). | |
609 | ||
610 | I realize that if I wanted to be true to the perl tradition, the only | |
611 | comment in this file would be something like: | |
612 | ||
613 | ...they shuffled back towards the rear of the line. 'No, not at the | |
614 | rear!' the slave-driver shouted. 'Three files up. And stay there... | |
615 | ||
616 | However, I really needed to violate that tradition just so I could keep | |
617 | track of what happens myself, not to mention some poor fool trying to | |
618 | understand this years from now :-). | |
619 | */ | |
620 | ||
621 | /* ********************************************************** Configuration */ | |
622 | ||
623 | #ifndef QSORT_ORDER_GUESS | |
624 | #define QSORT_ORDER_GUESS 2 /* Select doubling version of the netBSD trick */ | |
625 | #endif | |
626 | ||
627 | /* QSORT_MAX_STACK is the largest number of partitions that can be stacked up for | |
628 | future processing - a good max upper bound is log base 2 of memory size | |
629 | (32 on 32 bit machines, 64 on 64 bit machines, etc). In reality can | |
630 | safely be smaller than that since the program is taking up some space and | |
631 | most operating systems only let you grab some subset of contiguous | |
632 | memory (not to mention that you are normally sorting data larger than | |
633 | 1 byte element size :-). | |
634 | */ | |
635 | #ifndef QSORT_MAX_STACK | |
636 | #define QSORT_MAX_STACK 32 | |
637 | #endif | |
638 | ||
639 | /* QSORT_BREAK_EVEN is the size of the largest partition we should insertion sort. | |
640 | Anything bigger and we use qsort. If you make this too small, the qsort | |
641 | will probably break (or become less efficient), because it doesn't expect | |
642 | the middle element of a partition to be the same as the right or left - | |
643 | you have been warned). | |
644 | */ | |
645 | #ifndef QSORT_BREAK_EVEN | |
646 | #define QSORT_BREAK_EVEN 6 | |
647 | #endif | |
648 | ||
4eb872f6 JL |
649 | /* QSORT_PLAY_SAFE is the size of the largest partition we're willing |
650 | to go quadratic on. We innoculate larger partitions against | |
651 | quadratic behavior by shuffling them before sorting. This is not | |
652 | an absolute guarantee of non-quadratic behavior, but it would take | |
653 | staggeringly bad luck to pick extreme elements as the pivot | |
654 | from randomized data. | |
655 | */ | |
656 | #ifndef QSORT_PLAY_SAFE | |
657 | #define QSORT_PLAY_SAFE 255 | |
658 | #endif | |
659 | ||
84d4ea48 JH |
660 | /* ************************************************************* Data Types */ |
661 | ||
662 | /* hold left and right index values of a partition waiting to be sorted (the | |
663 | partition includes both left and right - right is NOT one past the end or | |
664 | anything like that). | |
665 | */ | |
666 | struct partition_stack_entry { | |
667 | int left; | |
668 | int right; | |
669 | #ifdef QSORT_ORDER_GUESS | |
670 | int qsort_break_even; | |
671 | #endif | |
672 | }; | |
673 | ||
674 | /* ******************************************************* Shorthand Macros */ | |
675 | ||
676 | /* Note that these macros will be used from inside the qsort function where | |
677 | we happen to know that the variable 'elt_size' contains the size of an | |
678 | array element and the variable 'temp' points to enough space to hold a | |
679 | temp element and the variable 'array' points to the array being sorted | |
680 | and 'compare' is the pointer to the compare routine. | |
681 | ||
682 | Also note that there are very many highly architecture specific ways | |
683 | these might be sped up, but this is simply the most generally portable | |
684 | code I could think of. | |
685 | */ | |
686 | ||
687 | /* Return < 0 == 0 or > 0 as the value of elt1 is < elt2, == elt2, > elt2 | |
688 | */ | |
689 | #define qsort_cmp(elt1, elt2) \ | |
690 | ((*compare)(aTHX_ array[elt1], array[elt2])) | |
691 | ||
692 | #ifdef QSORT_ORDER_GUESS | |
693 | #define QSORT_NOTICE_SWAP swapped++; | |
694 | #else | |
695 | #define QSORT_NOTICE_SWAP | |
696 | #endif | |
697 | ||
698 | /* swaps contents of array elements elt1, elt2. | |
699 | */ | |
700 | #define qsort_swap(elt1, elt2) \ | |
701 | STMT_START { \ | |
702 | QSORT_NOTICE_SWAP \ | |
703 | temp = array[elt1]; \ | |
704 | array[elt1] = array[elt2]; \ | |
705 | array[elt2] = temp; \ | |
706 | } STMT_END | |
707 | ||
708 | /* rotate contents of elt1, elt2, elt3 such that elt1 gets elt2, elt2 gets | |
709 | elt3 and elt3 gets elt1. | |
710 | */ | |
711 | #define qsort_rotate(elt1, elt2, elt3) \ | |
712 | STMT_START { \ | |
713 | QSORT_NOTICE_SWAP \ | |
714 | temp = array[elt1]; \ | |
715 | array[elt1] = array[elt2]; \ | |
716 | array[elt2] = array[elt3]; \ | |
717 | array[elt3] = temp; \ | |
718 | } STMT_END | |
719 | ||
720 | /* ************************************************************ Debug stuff */ | |
721 | ||
722 | #ifdef QSORT_DEBUG | |
723 | ||
724 | static void | |
725 | break_here() | |
726 | { | |
727 | return; /* good place to set a breakpoint */ | |
728 | } | |
729 | ||
730 | #define qsort_assert(t) (void)( (t) || (break_here(), 0) ) | |
731 | ||
732 | static void | |
733 | doqsort_all_asserts( | |
734 | void * array, | |
735 | size_t num_elts, | |
736 | size_t elt_size, | |
737 | int (*compare)(const void * elt1, const void * elt2), | |
738 | int pc_left, int pc_right, int u_left, int u_right) | |
739 | { | |
740 | int i; | |
741 | ||
742 | qsort_assert(pc_left <= pc_right); | |
743 | qsort_assert(u_right < pc_left); | |
744 | qsort_assert(pc_right < u_left); | |
745 | for (i = u_right + 1; i < pc_left; ++i) { | |
746 | qsort_assert(qsort_cmp(i, pc_left) < 0); | |
747 | } | |
748 | for (i = pc_left; i < pc_right; ++i) { | |
749 | qsort_assert(qsort_cmp(i, pc_right) == 0); | |
750 | } | |
751 | for (i = pc_right + 1; i < u_left; ++i) { | |
752 | qsort_assert(qsort_cmp(pc_right, i) < 0); | |
753 | } | |
754 | } | |
755 | ||
756 | #define qsort_all_asserts(PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT) \ | |
757 | doqsort_all_asserts(array, num_elts, elt_size, compare, \ | |
758 | PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT) | |
759 | ||
760 | #else | |
761 | ||
762 | #define qsort_assert(t) ((void)0) | |
763 | ||
764 | #define qsort_all_asserts(PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT) ((void)0) | |
765 | ||
766 | #endif | |
767 | ||
4eb872f6 | 768 | /* |
ccfc67b7 JH |
769 | =head1 Array Manipulation Functions |
770 | ||
84d4ea48 JH |
771 | =for apidoc sortsv |
772 | ||
8f5d5a51 | 773 | In-place sort an array of SV pointers with the given comparison routine. |
84d4ea48 | 774 | |
796b6530 | 775 | Currently this always uses mergesort. See C<L</sortsv_flags>> for a more |
7b9ef140 | 776 | flexible routine. |
78210658 | 777 | |
84d4ea48 JH |
778 | =cut |
779 | */ | |
4eb872f6 | 780 | |
84d4ea48 JH |
781 | void |
782 | Perl_sortsv(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp) | |
783 | { | |
7918f24d NC |
784 | PERL_ARGS_ASSERT_SORTSV; |
785 | ||
7b9ef140 | 786 | sortsv_flags(array, nmemb, cmp, 0); |
6c3fb703 NC |
787 | } |
788 | ||
4d562308 SF |
789 | #define SvNSIOK(sv) ((SvFLAGS(sv) & SVf_NOK) || ((SvFLAGS(sv) & (SVf_IOK|SVf_IVisUV)) == SVf_IOK)) |
790 | #define SvSIOK(sv) ((SvFLAGS(sv) & (SVf_IOK|SVf_IVisUV)) == SVf_IOK) | |
791 | #define SvNSIV(sv) ( SvNOK(sv) ? SvNVX(sv) : ( SvSIOK(sv) ? SvIVX(sv) : sv_2nv(sv) ) ) | |
792 | ||
84d4ea48 JH |
793 | PP(pp_sort) |
794 | { | |
20b7effb | 795 | dSP; dMARK; dORIGMARK; |
eb578fdb | 796 | SV **p1 = ORIGMARK+1, **p2; |
c70927a6 | 797 | SSize_t max, i; |
7d49f689 | 798 | AV* av = NULL; |
84d4ea48 | 799 | GV *gv; |
cbbf8932 | 800 | CV *cv = NULL; |
1c23e2bd | 801 | U8 gimme = GIMME_V; |
0bcc34c2 | 802 | OP* const nextop = PL_op->op_next; |
84d4ea48 JH |
803 | I32 overloading = 0; |
804 | bool hasargs = FALSE; | |
2b66f6d3 | 805 | bool copytmps; |
84d4ea48 | 806 | I32 is_xsub = 0; |
901017d6 AL |
807 | const U8 priv = PL_op->op_private; |
808 | const U8 flags = PL_op->op_flags; | |
7b9ef140 RH |
809 | U32 sort_flags = 0; |
810 | void (*sortsvp)(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp, U32 flags) | |
811 | = Perl_sortsv_flags; | |
4d562308 | 812 | I32 all_SIVs = 1; |
84d4ea48 | 813 | |
7b9ef140 RH |
814 | if ((priv & OPpSORT_DESCEND) != 0) |
815 | sort_flags |= SORTf_DESC; | |
7b9ef140 RH |
816 | if ((priv & OPpSORT_STABLE) != 0) |
817 | sort_flags |= SORTf_STABLE; | |
afe59f35 FC |
818 | if ((priv & OPpSORT_UNSTABLE) != 0) |
819 | sort_flags |= SORTf_UNSTABLE; | |
7b9ef140 | 820 | |
84d4ea48 JH |
821 | if (gimme != G_ARRAY) { |
822 | SP = MARK; | |
b59aed67 | 823 | EXTEND(SP,1); |
84d4ea48 JH |
824 | RETPUSHUNDEF; |
825 | } | |
826 | ||
827 | ENTER; | |
828 | SAVEVPTR(PL_sortcop); | |
471178c0 NC |
829 | if (flags & OPf_STACKED) { |
830 | if (flags & OPf_SPECIAL) { | |
e6dae479 | 831 | OP *nullop = OpSIBLING(cLISTOP->op_first); /* pass pushmark */ |
932bca29 DM |
832 | assert(nullop->op_type == OP_NULL); |
833 | PL_sortcop = nullop->op_next; | |
84d4ea48 JH |
834 | } |
835 | else { | |
f7bc00ea | 836 | GV *autogv = NULL; |
5a34f1cd | 837 | HV *stash; |
f7bc00ea FC |
838 | cv = sv_2cv(*++MARK, &stash, &gv, GV_ADD); |
839 | check_cv: | |
84d4ea48 | 840 | if (cv && SvPOK(cv)) { |
ad64d0ec | 841 | const char * const proto = SvPV_nolen_const(MUTABLE_SV(cv)); |
84d4ea48 JH |
842 | if (proto && strEQ(proto, "$$")) { |
843 | hasargs = TRUE; | |
844 | } | |
845 | } | |
2fc49ef1 FC |
846 | if (cv && CvISXSUB(cv) && CvXSUB(cv)) { |
847 | is_xsub = 1; | |
848 | } | |
849 | else if (!(cv && CvROOT(cv))) { | |
850 | if (gv) { | |
f7bc00ea FC |
851 | goto autoload; |
852 | } | |
853 | else if (!CvANON(cv) && (gv = CvGV(cv))) { | |
854 | if (cv != GvCV(gv)) cv = GvCV(gv); | |
855 | autoload: | |
856 | if (!autogv && ( | |
857 | autogv = gv_autoload_pvn( | |
858 | GvSTASH(gv), GvNAME(gv), GvNAMELEN(gv), | |
859 | GvNAMEUTF8(gv) ? SVf_UTF8 : 0 | |
860 | ) | |
861 | )) { | |
862 | cv = GvCVu(autogv); | |
863 | goto check_cv; | |
864 | } | |
865 | else { | |
84d4ea48 | 866 | SV *tmpstr = sv_newmortal(); |
bd61b366 | 867 | gv_efullname3(tmpstr, gv, NULL); |
147e3846 | 868 | DIE(aTHX_ "Undefined sort subroutine \"%" SVf "\" called", |
be2597df | 869 | SVfARG(tmpstr)); |
f7bc00ea | 870 | } |
84d4ea48 JH |
871 | } |
872 | else { | |
873 | DIE(aTHX_ "Undefined subroutine in sort"); | |
874 | } | |
875 | } | |
876 | ||
877 | if (is_xsub) | |
878 | PL_sortcop = (OP*)cv; | |
9850bf21 | 879 | else |
84d4ea48 | 880 | PL_sortcop = CvSTART(cv); |
84d4ea48 JH |
881 | } |
882 | } | |
883 | else { | |
5f66b61c | 884 | PL_sortcop = NULL; |
84d4ea48 JH |
885 | } |
886 | ||
84721d61 DM |
887 | /* optimiser converts "@a = sort @a" to "sort \@a". In this case, |
888 | * push (@a) onto stack, then assign result back to @a at the end of | |
889 | * this function */ | |
0723351e | 890 | if (priv & OPpSORT_INPLACE) { |
fe1bc4cf DM |
891 | assert( MARK+1 == SP && *SP && SvTYPE(*SP) == SVt_PVAV); |
892 | (void)POPMARK; /* remove mark associated with ex-OP_AASSIGN */ | |
502c6561 | 893 | av = MUTABLE_AV((*SP)); |
84721d61 DM |
894 | if (SvREADONLY(av)) |
895 | Perl_croak_no_modify(); | |
fe1bc4cf | 896 | max = AvFILL(av) + 1; |
84721d61 | 897 | MEXTEND(SP, max); |
fe1bc4cf | 898 | if (SvMAGICAL(av)) { |
fe2774ed | 899 | for (i=0; i < max; i++) { |
fe1bc4cf | 900 | SV **svp = av_fetch(av, i, FALSE); |
a0714e2c | 901 | *SP++ = (svp) ? *svp : NULL; |
fe1bc4cf DM |
902 | } |
903 | } | |
84721d61 DM |
904 | else { |
905 | SV **svp = AvARRAY(av); | |
906 | assert(svp || max == 0); | |
907 | for (i = 0; i < max; i++) | |
908 | *SP++ = *svp++; | |
fe1bc4cf | 909 | } |
84721d61 DM |
910 | SP--; |
911 | p1 = p2 = SP - (max-1); | |
fe1bc4cf DM |
912 | } |
913 | else { | |
914 | p2 = MARK+1; | |
915 | max = SP - MARK; | |
916 | } | |
917 | ||
83a44efe SF |
918 | /* shuffle stack down, removing optional initial cv (p1!=p2), plus |
919 | * any nulls; also stringify or converting to integer or number as | |
920 | * required any args */ | |
ff859a7f | 921 | copytmps = cBOOL(PL_sortcop); |
fe1bc4cf DM |
922 | for (i=max; i > 0 ; i--) { |
923 | if ((*p1 = *p2++)) { /* Weed out nulls. */ | |
60779a30 | 924 | if (copytmps && SvPADTMP(*p1)) { |
2b66f6d3 | 925 | *p1 = sv_mortalcopy(*p1); |
60779a30 | 926 | } |
fe1bc4cf | 927 | SvTEMP_off(*p1); |
83a44efe SF |
928 | if (!PL_sortcop) { |
929 | if (priv & OPpSORT_NUMERIC) { | |
930 | if (priv & OPpSORT_INTEGER) { | |
bdbefedf DM |
931 | if (!SvIOK(*p1)) |
932 | (void)sv_2iv_flags(*p1, SV_GMAGIC|SV_SKIP_OVERLOAD); | |
83a44efe SF |
933 | } |
934 | else { | |
bdbefedf DM |
935 | if (!SvNSIOK(*p1)) |
936 | (void)sv_2nv_flags(*p1, SV_GMAGIC|SV_SKIP_OVERLOAD); | |
4d562308 SF |
937 | if (all_SIVs && !SvSIOK(*p1)) |
938 | all_SIVs = 0; | |
83a44efe SF |
939 | } |
940 | } | |
941 | else { | |
bdbefedf DM |
942 | if (!SvPOK(*p1)) |
943 | (void)sv_2pv_flags(*p1, 0, | |
944 | SV_GMAGIC|SV_CONST_RETURN|SV_SKIP_OVERLOAD); | |
83a44efe | 945 | } |
bdbefedf DM |
946 | if (SvAMAGIC(*p1)) |
947 | overloading = 1; | |
84d4ea48 | 948 | } |
fe1bc4cf | 949 | p1++; |
84d4ea48 | 950 | } |
fe1bc4cf DM |
951 | else |
952 | max--; | |
84d4ea48 | 953 | } |
fe1bc4cf | 954 | if (max > 1) { |
471178c0 | 955 | SV **start; |
fe1bc4cf | 956 | if (PL_sortcop) { |
84d4ea48 | 957 | PERL_CONTEXT *cx; |
901017d6 | 958 | const bool oldcatch = CATCH_GET; |
8ae997c5 | 959 | I32 old_savestack_ix = PL_savestack_ix; |
84d4ea48 | 960 | |
84d4ea48 JH |
961 | SAVEOP(); |
962 | ||
963 | CATCH_SET(TRUE); | |
964 | PUSHSTACKi(PERLSI_SORT); | |
965 | if (!hasargs && !is_xsub) { | |
8465ba45 FC |
966 | SAVEGENERICSV(PL_firstgv); |
967 | SAVEGENERICSV(PL_secondgv); | |
968 | PL_firstgv = MUTABLE_GV(SvREFCNT_inc( | |
969 | gv_fetchpvs("a", GV_ADD|GV_NOTQUAL, SVt_PV) | |
970 | )); | |
971 | PL_secondgv = MUTABLE_GV(SvREFCNT_inc( | |
972 | gv_fetchpvs("b", GV_ADD|GV_NOTQUAL, SVt_PV) | |
973 | )); | |
dc9ef998 TC |
974 | /* make sure the GP isn't removed out from under us for |
975 | * the SAVESPTR() */ | |
976 | save_gp(PL_firstgv, 0); | |
977 | save_gp(PL_secondgv, 0); | |
978 | /* we don't want modifications localized */ | |
979 | GvINTRO_off(PL_firstgv); | |
980 | GvINTRO_off(PL_secondgv); | |
16ada235 Z |
981 | SAVEGENERICSV(GvSV(PL_firstgv)); |
982 | SvREFCNT_inc(GvSV(PL_firstgv)); | |
983 | SAVEGENERICSV(GvSV(PL_secondgv)); | |
984 | SvREFCNT_inc(GvSV(PL_secondgv)); | |
84d4ea48 JH |
985 | } |
986 | ||
33411212 | 987 | gimme = G_SCALAR; |
ed8ff0f3 | 988 | cx = cx_pushblock(CXt_NULL, gimme, PL_stack_base, old_savestack_ix); |
471178c0 | 989 | if (!(flags & OPf_SPECIAL)) { |
79646418 | 990 | cx->cx_type = CXt_SUB|CXp_MULTICALL; |
a73d8813 | 991 | cx_pushsub(cx, cv, NULL, hasargs); |
9850bf21 | 992 | if (!is_xsub) { |
b70d5558 | 993 | PADLIST * const padlist = CvPADLIST(cv); |
9850bf21 | 994 | |
d2af2719 | 995 | if (++CvDEPTH(cv) >= 2) |
9850bf21 | 996 | pad_push(padlist, CvDEPTH(cv)); |
9850bf21 | 997 | PAD_SET_CUR_NOSAVE(padlist, CvDEPTH(cv)); |
84d4ea48 | 998 | |
9850bf21 RH |
999 | if (hasargs) { |
1000 | /* This is mostly copied from pp_entersub */ | |
502c6561 | 1001 | AV * const av = MUTABLE_AV(PAD_SVl(0)); |
84d4ea48 | 1002 | |
9850bf21 | 1003 | cx->blk_sub.savearray = GvAV(PL_defgv); |
502c6561 | 1004 | GvAV(PL_defgv) = MUTABLE_AV(SvREFCNT_inc_simple(av)); |
9850bf21 RH |
1005 | } |
1006 | ||
1007 | } | |
84d4ea48 | 1008 | } |
486430a5 | 1009 | |
471178c0 NC |
1010 | start = p1 - max; |
1011 | sortsvp(aTHX_ start, max, | |
7b9ef140 RH |
1012 | (is_xsub ? S_sortcv_xsub : hasargs ? S_sortcv_stacked : S_sortcv), |
1013 | sort_flags); | |
84d4ea48 | 1014 | |
4df352a8 | 1015 | /* Reset cx, in case the context stack has been reallocated. */ |
4ebe6e95 | 1016 | cx = CX_CUR(); |
4df352a8 DM |
1017 | |
1018 | PL_stack_sp = PL_stack_base + cx->blk_oldsp; | |
1019 | ||
2f450c1b | 1020 | CX_LEAVE_SCOPE(cx); |
9850bf21 | 1021 | if (!(flags & OPf_SPECIAL)) { |
4df352a8 | 1022 | assert(CxTYPE(cx) == CXt_SUB); |
a73d8813 | 1023 | cx_popsub(cx); |
9850bf21 | 1024 | } |
2f450c1b | 1025 | else |
4df352a8 | 1026 | assert(CxTYPE(cx) == CXt_NULL); |
2f450c1b | 1027 | /* there isn't a POPNULL ! */ |
1dfbe6b4 | 1028 | |
ed8ff0f3 | 1029 | cx_popblock(cx); |
5da525e9 | 1030 | CX_POP(cx); |
84d4ea48 JH |
1031 | POPSTACK; |
1032 | CATCH_SET(oldcatch); | |
1033 | } | |
fe1bc4cf | 1034 | else { |
84d4ea48 | 1035 | MEXTEND(SP, 20); /* Can't afford stack realloc on signal. */ |
84721d61 | 1036 | start = ORIGMARK+1; |
471178c0 | 1037 | sortsvp(aTHX_ start, max, |
0723351e | 1038 | (priv & OPpSORT_NUMERIC) |
4d562308 | 1039 | ? ( ( ( priv & OPpSORT_INTEGER) || all_SIVs) |
f0f5dc9d AL |
1040 | ? ( overloading ? S_amagic_i_ncmp : S_sv_i_ncmp) |
1041 | : ( overloading ? S_amagic_ncmp : S_sv_ncmp ) ) | |
130c5df3 KW |
1042 | : ( |
1043 | #ifdef USE_LOCALE_COLLATE | |
1044 | IN_LC_RUNTIME(LC_COLLATE) | |
84d4ea48 | 1045 | ? ( overloading |
d3fcec1f SP |
1046 | ? (SVCOMPARE_t)S_amagic_cmp_locale |
1047 | : (SVCOMPARE_t)sv_cmp_locale_static) | |
130c5df3 KW |
1048 | : |
1049 | #endif | |
1050 | ( overloading ? (SVCOMPARE_t)S_amagic_cmp : (SVCOMPARE_t)sv_cmp_static)), | |
7b9ef140 | 1051 | sort_flags); |
471178c0 | 1052 | } |
7b9ef140 | 1053 | if ((priv & OPpSORT_REVERSE) != 0) { |
471178c0 NC |
1054 | SV **q = start+max-1; |
1055 | while (start < q) { | |
0bcc34c2 | 1056 | SV * const tmp = *start; |
471178c0 NC |
1057 | *start++ = *q; |
1058 | *q-- = tmp; | |
84d4ea48 JH |
1059 | } |
1060 | } | |
1061 | } | |
84721d61 DM |
1062 | |
1063 | if (av) { | |
1064 | /* copy back result to the array */ | |
1065 | SV** const base = MARK+1; | |
1066 | if (SvMAGICAL(av)) { | |
1067 | for (i = 0; i < max; i++) | |
1068 | base[i] = newSVsv(base[i]); | |
1069 | av_clear(av); | |
1070 | av_extend(av, max); | |
1071 | for (i=0; i < max; i++) { | |
1072 | SV * const sv = base[i]; | |
1073 | SV ** const didstore = av_store(av, i, sv); | |
1074 | if (SvSMAGICAL(sv)) | |
1075 | mg_set(sv); | |
1076 | if (!didstore) | |
1077 | sv_2mortal(sv); | |
1078 | } | |
1079 | } | |
1080 | else { | |
1081 | /* the elements of av are likely to be the same as the | |
1082 | * (non-refcounted) elements on the stack, just in a different | |
1083 | * order. However, its possible that someone's messed with av | |
1084 | * in the meantime. So bump and unbump the relevant refcounts | |
1085 | * first. | |
1086 | */ | |
45c198c1 DM |
1087 | for (i = 0; i < max; i++) { |
1088 | SV *sv = base[i]; | |
1089 | assert(sv); | |
1090 | if (SvREFCNT(sv) > 1) | |
1091 | base[i] = newSVsv(sv); | |
1092 | else | |
1093 | SvREFCNT_inc_simple_void_NN(sv); | |
1094 | } | |
84721d61 DM |
1095 | av_clear(av); |
1096 | if (max > 0) { | |
1097 | av_extend(av, max); | |
1098 | Copy(base, AvARRAY(av), max, SV*); | |
1099 | } | |
1100 | AvFILLp(av) = max - 1; | |
1101 | AvREIFY_off(av); | |
1102 | AvREAL_on(av); | |
1103 | } | |
fe1bc4cf | 1104 | } |
84d4ea48 | 1105 | LEAVE; |
84721d61 | 1106 | PL_stack_sp = ORIGMARK + max; |
84d4ea48 JH |
1107 | return nextop; |
1108 | } | |
1109 | ||
1110 | static I32 | |
31e9e0a3 | 1111 | S_sortcv(pTHX_ SV *const a, SV *const b) |
84d4ea48 | 1112 | { |
901017d6 | 1113 | const I32 oldsaveix = PL_savestack_ix; |
84d4ea48 | 1114 | I32 result; |
ad021bfb | 1115 | PMOP * const pm = PL_curpm; |
a9ea019a | 1116 | COP * const cop = PL_curcop; |
16ada235 | 1117 | SV *olda, *oldb; |
7918f24d NC |
1118 | |
1119 | PERL_ARGS_ASSERT_SORTCV; | |
1120 | ||
16ada235 Z |
1121 | olda = GvSV(PL_firstgv); |
1122 | GvSV(PL_firstgv) = SvREFCNT_inc_simple_NN(a); | |
1123 | SvREFCNT_dec(olda); | |
1124 | oldb = GvSV(PL_secondgv); | |
1125 | GvSV(PL_secondgv) = SvREFCNT_inc_simple_NN(b); | |
1126 | SvREFCNT_dec(oldb); | |
84d4ea48 JH |
1127 | PL_stack_sp = PL_stack_base; |
1128 | PL_op = PL_sortcop; | |
1129 | CALLRUNOPS(aTHX); | |
a9ea019a | 1130 | PL_curcop = cop; |
33411212 DM |
1131 | /* entry zero of a stack is always PL_sv_undef, which |
1132 | * simplifies converting a '()' return into undef in scalar context */ | |
1133 | assert(PL_stack_sp > PL_stack_base || *PL_stack_base == &PL_sv_undef); | |
1134 | result = SvIV(*PL_stack_sp); | |
626ed49c | 1135 | |
53d3542d | 1136 | LEAVE_SCOPE(oldsaveix); |
ad021bfb | 1137 | PL_curpm = pm; |
84d4ea48 JH |
1138 | return result; |
1139 | } | |
1140 | ||
1141 | static I32 | |
31e9e0a3 | 1142 | S_sortcv_stacked(pTHX_ SV *const a, SV *const b) |
84d4ea48 | 1143 | { |
901017d6 | 1144 | const I32 oldsaveix = PL_savestack_ix; |
84d4ea48 | 1145 | I32 result; |
901017d6 | 1146 | AV * const av = GvAV(PL_defgv); |
ad021bfb | 1147 | PMOP * const pm = PL_curpm; |
a9ea019a | 1148 | COP * const cop = PL_curcop; |
84d4ea48 | 1149 | |
7918f24d NC |
1150 | PERL_ARGS_ASSERT_SORTCV_STACKED; |
1151 | ||
8f443ca6 GG |
1152 | if (AvREAL(av)) { |
1153 | av_clear(av); | |
1154 | AvREAL_off(av); | |
1155 | AvREIFY_on(av); | |
1156 | } | |
84d4ea48 | 1157 | if (AvMAX(av) < 1) { |
8f443ca6 | 1158 | SV **ary = AvALLOC(av); |
84d4ea48 JH |
1159 | if (AvARRAY(av) != ary) { |
1160 | AvMAX(av) += AvARRAY(av) - AvALLOC(av); | |
9c6bc640 | 1161 | AvARRAY(av) = ary; |
84d4ea48 JH |
1162 | } |
1163 | if (AvMAX(av) < 1) { | |
84d4ea48 | 1164 | Renew(ary,2,SV*); |
00195859 | 1165 | AvMAX(av) = 1; |
9c6bc640 | 1166 | AvARRAY(av) = ary; |
8f443ca6 | 1167 | AvALLOC(av) = ary; |
84d4ea48 JH |
1168 | } |
1169 | } | |
1170 | AvFILLp(av) = 1; | |
1171 | ||
1172 | AvARRAY(av)[0] = a; | |
1173 | AvARRAY(av)[1] = b; | |
1174 | PL_stack_sp = PL_stack_base; | |
1175 | PL_op = PL_sortcop; | |
1176 | CALLRUNOPS(aTHX); | |
a9ea019a | 1177 | PL_curcop = cop; |
33411212 DM |
1178 | /* entry zero of a stack is always PL_sv_undef, which |
1179 | * simplifies converting a '()' return into undef in scalar context */ | |
1180 | assert(PL_stack_sp > PL_stack_base || *PL_stack_base == &PL_sv_undef); | |
1181 | result = SvIV(*PL_stack_sp); | |
626ed49c | 1182 | |
53d3542d | 1183 | LEAVE_SCOPE(oldsaveix); |
ad021bfb | 1184 | PL_curpm = pm; |
84d4ea48 JH |
1185 | return result; |
1186 | } | |
1187 | ||
1188 | static I32 | |
31e9e0a3 | 1189 | S_sortcv_xsub(pTHX_ SV *const a, SV *const b) |
84d4ea48 | 1190 | { |
20b7effb | 1191 | dSP; |
901017d6 | 1192 | const I32 oldsaveix = PL_savestack_ix; |
ea726b52 | 1193 | CV * const cv=MUTABLE_CV(PL_sortcop); |
84d4ea48 | 1194 | I32 result; |
ad021bfb | 1195 | PMOP * const pm = PL_curpm; |
84d4ea48 | 1196 | |
7918f24d NC |
1197 | PERL_ARGS_ASSERT_SORTCV_XSUB; |
1198 | ||
84d4ea48 JH |
1199 | SP = PL_stack_base; |
1200 | PUSHMARK(SP); | |
1201 | EXTEND(SP, 2); | |
1202 | *++SP = a; | |
1203 | *++SP = b; | |
1204 | PUTBACK; | |
1205 | (void)(*CvXSUB(cv))(aTHX_ cv); | |
33411212 DM |
1206 | /* entry zero of a stack is always PL_sv_undef, which |
1207 | * simplifies converting a '()' return into undef in scalar context */ | |
1208 | assert(PL_stack_sp > PL_stack_base || *PL_stack_base == &PL_sv_undef); | |
84d4ea48 | 1209 | result = SvIV(*PL_stack_sp); |
33411212 | 1210 | |
53d3542d | 1211 | LEAVE_SCOPE(oldsaveix); |
ad021bfb | 1212 | PL_curpm = pm; |
84d4ea48 JH |
1213 | return result; |
1214 | } | |
1215 | ||
1216 | ||
1217 | static I32 | |
31e9e0a3 | 1218 | S_sv_ncmp(pTHX_ SV *const a, SV *const b) |
84d4ea48 | 1219 | { |
427fbfe8 | 1220 | I32 cmp = do_ncmp(a, b); |
7918f24d NC |
1221 | |
1222 | PERL_ARGS_ASSERT_SV_NCMP; | |
1223 | ||
427fbfe8 | 1224 | if (cmp == 2) { |
f3dab52a FC |
1225 | if (ckWARN(WARN_UNINITIALIZED)) report_uninit(NULL); |
1226 | return 0; | |
1227 | } | |
427fbfe8 TC |
1228 | |
1229 | return cmp; | |
84d4ea48 JH |
1230 | } |
1231 | ||
1232 | static I32 | |
31e9e0a3 | 1233 | S_sv_i_ncmp(pTHX_ SV *const a, SV *const b) |
84d4ea48 | 1234 | { |
901017d6 AL |
1235 | const IV iv1 = SvIV(a); |
1236 | const IV iv2 = SvIV(b); | |
7918f24d NC |
1237 | |
1238 | PERL_ARGS_ASSERT_SV_I_NCMP; | |
1239 | ||
84d4ea48 JH |
1240 | return iv1 < iv2 ? -1 : iv1 > iv2 ? 1 : 0; |
1241 | } | |
901017d6 AL |
1242 | |
1243 | #define tryCALL_AMAGICbin(left,right,meth) \ | |
79a8d529 | 1244 | (SvAMAGIC(left)||SvAMAGIC(right)) \ |
31d632c3 | 1245 | ? amagic_call(left, right, meth, 0) \ |
a0714e2c | 1246 | : NULL; |
84d4ea48 | 1247 | |
659c4b96 | 1248 | #define SORT_NORMAL_RETURN_VALUE(val) (((val) > 0) ? 1 : ((val) ? -1 : 0)) |
eeb9de02 | 1249 | |
84d4ea48 | 1250 | static I32 |
5aaab254 | 1251 | S_amagic_ncmp(pTHX_ SV *const a, SV *const b) |
84d4ea48 | 1252 | { |
31d632c3 | 1253 | SV * const tmpsv = tryCALL_AMAGICbin(a,b,ncmp_amg); |
7918f24d NC |
1254 | |
1255 | PERL_ARGS_ASSERT_AMAGIC_NCMP; | |
1256 | ||
84d4ea48 | 1257 | if (tmpsv) { |
84d4ea48 | 1258 | if (SvIOK(tmpsv)) { |
901017d6 | 1259 | const I32 i = SvIVX(tmpsv); |
eeb9de02 | 1260 | return SORT_NORMAL_RETURN_VALUE(i); |
84d4ea48 | 1261 | } |
901017d6 AL |
1262 | else { |
1263 | const NV d = SvNV(tmpsv); | |
eeb9de02 | 1264 | return SORT_NORMAL_RETURN_VALUE(d); |
901017d6 | 1265 | } |
84d4ea48 | 1266 | } |
f0f5dc9d | 1267 | return S_sv_ncmp(aTHX_ a, b); |
84d4ea48 JH |
1268 | } |
1269 | ||
1270 | static I32 | |
5aaab254 | 1271 | S_amagic_i_ncmp(pTHX_ SV *const a, SV *const b) |
84d4ea48 | 1272 | { |
31d632c3 | 1273 | SV * const tmpsv = tryCALL_AMAGICbin(a,b,ncmp_amg); |
7918f24d NC |
1274 | |
1275 | PERL_ARGS_ASSERT_AMAGIC_I_NCMP; | |
1276 | ||
84d4ea48 | 1277 | if (tmpsv) { |
84d4ea48 | 1278 | if (SvIOK(tmpsv)) { |
901017d6 | 1279 | const I32 i = SvIVX(tmpsv); |
eeb9de02 | 1280 | return SORT_NORMAL_RETURN_VALUE(i); |
84d4ea48 | 1281 | } |
901017d6 AL |
1282 | else { |
1283 | const NV d = SvNV(tmpsv); | |
eeb9de02 | 1284 | return SORT_NORMAL_RETURN_VALUE(d); |
901017d6 | 1285 | } |
84d4ea48 | 1286 | } |
f0f5dc9d | 1287 | return S_sv_i_ncmp(aTHX_ a, b); |
84d4ea48 JH |
1288 | } |
1289 | ||
1290 | static I32 | |
5aaab254 | 1291 | S_amagic_cmp(pTHX_ SV *const str1, SV *const str2) |
84d4ea48 | 1292 | { |
31d632c3 | 1293 | SV * const tmpsv = tryCALL_AMAGICbin(str1,str2,scmp_amg); |
7918f24d NC |
1294 | |
1295 | PERL_ARGS_ASSERT_AMAGIC_CMP; | |
1296 | ||
84d4ea48 | 1297 | if (tmpsv) { |
84d4ea48 | 1298 | if (SvIOK(tmpsv)) { |
901017d6 | 1299 | const I32 i = SvIVX(tmpsv); |
eeb9de02 | 1300 | return SORT_NORMAL_RETURN_VALUE(i); |
84d4ea48 | 1301 | } |
901017d6 AL |
1302 | else { |
1303 | const NV d = SvNV(tmpsv); | |
eeb9de02 | 1304 | return SORT_NORMAL_RETURN_VALUE(d); |
901017d6 | 1305 | } |
84d4ea48 JH |
1306 | } |
1307 | return sv_cmp(str1, str2); | |
1308 | } | |
1309 | ||
91191cf7 KW |
1310 | #ifdef USE_LOCALE_COLLATE |
1311 | ||
84d4ea48 | 1312 | static I32 |
5aaab254 | 1313 | S_amagic_cmp_locale(pTHX_ SV *const str1, SV *const str2) |
84d4ea48 | 1314 | { |
31d632c3 | 1315 | SV * const tmpsv = tryCALL_AMAGICbin(str1,str2,scmp_amg); |
7918f24d NC |
1316 | |
1317 | PERL_ARGS_ASSERT_AMAGIC_CMP_LOCALE; | |
1318 | ||
84d4ea48 | 1319 | if (tmpsv) { |
84d4ea48 | 1320 | if (SvIOK(tmpsv)) { |
901017d6 | 1321 | const I32 i = SvIVX(tmpsv); |
eeb9de02 | 1322 | return SORT_NORMAL_RETURN_VALUE(i); |
84d4ea48 | 1323 | } |
901017d6 AL |
1324 | else { |
1325 | const NV d = SvNV(tmpsv); | |
eeb9de02 | 1326 | return SORT_NORMAL_RETURN_VALUE(d); |
901017d6 | 1327 | } |
84d4ea48 JH |
1328 | } |
1329 | return sv_cmp_locale(str1, str2); | |
1330 | } | |
241d1a3b | 1331 | |
91191cf7 KW |
1332 | #endif |
1333 | ||
241d1a3b | 1334 | /* |
14d04a33 | 1335 | * ex: set ts=8 sts=4 sw=4 et: |
37442d52 | 1336 | */ |