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 | ||
42165d27 VK |
33 | #if defined(UNDER_CE) |
34 | /* looks like 'small' is reserved word for WINCE (or somesuch)*/ | |
35 | #define small xsmall | |
36 | #endif | |
37 | ||
84d4ea48 JH |
38 | #define sv_cmp_static Perl_sv_cmp |
39 | #define sv_cmp_locale_static Perl_sv_cmp_locale | |
40 | ||
c53fc8a6 JH |
41 | #ifndef SMALLSORT |
42 | #define SMALLSORT (200) | |
43 | #endif | |
44 | ||
7b9ef140 RH |
45 | /* Flags for qsortsv and mergesortsv */ |
46 | #define SORTf_DESC 1 | |
47 | #define SORTf_STABLE 2 | |
afe59f35 | 48 | #define SORTf_UNSTABLE 8 |
7b9ef140 | 49 | |
84d4ea48 JH |
50 | /* |
51 | * The mergesort implementation is by Peter M. Mcilroy <pmcilroy@lucent.com>. | |
52 | * | |
53 | * The original code was written in conjunction with BSD Computer Software | |
54 | * Research Group at University of California, Berkeley. | |
55 | * | |
393db44d JL |
56 | * See also: "Optimistic Sorting and Information Theoretic Complexity" |
57 | * Peter McIlroy | |
58 | * SODA (Fourth Annual ACM-SIAM Symposium on Discrete Algorithms), | |
59 | * pp 467-474, Austin, Texas, 25-27 January 1993. | |
84d4ea48 | 60 | * |
393db44d | 61 | * The integration to Perl is by John P. Linderman <jpl.jpl@gmail.com>. |
84d4ea48 JH |
62 | * |
63 | * The code can be distributed under the same terms as Perl itself. | |
64 | * | |
65 | */ | |
66 | ||
84d4ea48 JH |
67 | |
68 | typedef char * aptr; /* pointer for arithmetic on sizes */ | |
69 | typedef SV * gptr; /* pointers in our lists */ | |
70 | ||
71 | /* Binary merge internal sort, with a few special mods | |
72 | ** for the special perl environment it now finds itself in. | |
73 | ** | |
74 | ** Things that were once options have been hotwired | |
75 | ** to values suitable for this use. In particular, we'll always | |
76 | ** initialize looking for natural runs, we'll always produce stable | |
77 | ** output, and we'll always do Peter McIlroy's binary merge. | |
78 | */ | |
79 | ||
80 | /* Pointer types for arithmetic and storage and convenience casts */ | |
81 | ||
82 | #define APTR(P) ((aptr)(P)) | |
83 | #define GPTP(P) ((gptr *)(P)) | |
84 | #define GPPP(P) ((gptr **)(P)) | |
85 | ||
86 | ||
87 | /* byte offset from pointer P to (larger) pointer Q */ | |
88 | #define BYTEOFF(P, Q) (APTR(Q) - APTR(P)) | |
89 | ||
90 | #define PSIZE sizeof(gptr) | |
91 | ||
92 | /* If PSIZE is power of 2, make PSHIFT that power, if that helps */ | |
93 | ||
94 | #ifdef PSHIFT | |
95 | #define PNELEM(P, Q) (BYTEOFF(P,Q) >> (PSHIFT)) | |
96 | #define PNBYTE(N) ((N) << (PSHIFT)) | |
97 | #define PINDEX(P, N) (GPTP(APTR(P) + PNBYTE(N))) | |
98 | #else | |
99 | /* Leave optimization to compiler */ | |
100 | #define PNELEM(P, Q) (GPTP(Q) - GPTP(P)) | |
101 | #define PNBYTE(N) ((N) * (PSIZE)) | |
102 | #define PINDEX(P, N) (GPTP(P) + (N)) | |
103 | #endif | |
104 | ||
105 | /* Pointer into other corresponding to pointer into this */ | |
106 | #define POTHER(P, THIS, OTHER) GPTP(APTR(OTHER) + BYTEOFF(THIS,P)) | |
107 | ||
108 | #define FROMTOUPTO(src, dst, lim) do *dst++ = *src++; while(src<lim) | |
109 | ||
110 | ||
486ec47a | 111 | /* Runs are identified by a pointer in the auxiliary list. |
84d4ea48 JH |
112 | ** The pointer is at the start of the list, |
113 | ** and it points to the start of the next list. | |
114 | ** NEXT is used as an lvalue, too. | |
115 | */ | |
116 | ||
117 | #define NEXT(P) (*GPPP(P)) | |
118 | ||
119 | ||
120 | /* PTHRESH is the minimum number of pairs with the same sense to justify | |
121 | ** checking for a run and extending it. Note that PTHRESH counts PAIRS, | |
122 | ** not just elements, so PTHRESH == 8 means a run of 16. | |
123 | */ | |
124 | ||
125 | #define PTHRESH (8) | |
126 | ||
127 | /* RTHRESH is the number of elements in a run that must compare low | |
128 | ** to the low element from the opposing run before we justify | |
129 | ** doing a binary rampup instead of single stepping. | |
130 | ** In random input, N in a row low should only happen with | |
131 | ** probability 2^(1-N), so we can risk that we are dealing | |
132 | ** with orderly input without paying much when we aren't. | |
133 | */ | |
134 | ||
135 | #define RTHRESH (6) | |
136 | ||
137 | ||
138 | /* | |
139 | ** Overview of algorithm and variables. | |
140 | ** The array of elements at list1 will be organized into runs of length 2, | |
141 | ** or runs of length >= 2 * PTHRESH. We only try to form long runs when | |
142 | ** PTHRESH adjacent pairs compare in the same way, suggesting overall order. | |
143 | ** | |
144 | ** Unless otherwise specified, pair pointers address the first of two elements. | |
145 | ** | |
a0288114 AL |
146 | ** b and b+1 are a pair that compare with sense "sense". |
147 | ** b is the "bottom" of adjacent pairs that might form a longer run. | |
84d4ea48 JH |
148 | ** |
149 | ** p2 parallels b in the list2 array, where runs are defined by | |
150 | ** a pointer chain. | |
151 | ** | |
a0288114 | 152 | ** t represents the "top" of the adjacent pairs that might extend |
84d4ea48 JH |
153 | ** the run beginning at b. Usually, t addresses a pair |
154 | ** that compares with opposite sense from (b,b+1). | |
155 | ** However, it may also address a singleton element at the end of list1, | |
a0288114 | 156 | ** or it may be equal to "last", the first element beyond list1. |
84d4ea48 JH |
157 | ** |
158 | ** r addresses the Nth pair following b. If this would be beyond t, | |
159 | ** we back it off to t. Only when r is less than t do we consider the | |
160 | ** run long enough to consider checking. | |
161 | ** | |
162 | ** q addresses a pair such that the pairs at b through q already form a run. | |
163 | ** Often, q will equal b, indicating we only are sure of the pair itself. | |
164 | ** However, a search on the previous cycle may have revealed a longer run, | |
165 | ** so q may be greater than b. | |
166 | ** | |
167 | ** p is used to work back from a candidate r, trying to reach q, | |
168 | ** which would mean b through r would be a run. If we discover such a run, | |
169 | ** we start q at r and try to push it further towards t. | |
170 | ** If b through r is NOT a run, we detect the wrong order at (p-1,p). | |
171 | ** In any event, after the check (if any), we have two main cases. | |
172 | ** | |
173 | ** 1) Short run. b <= q < p <= r <= t. | |
174 | ** b through q is a run (perhaps trivial) | |
175 | ** q through p are uninteresting pairs | |
176 | ** p through r is a run | |
177 | ** | |
178 | ** 2) Long run. b < r <= q < t. | |
179 | ** b through q is a run (of length >= 2 * PTHRESH) | |
180 | ** | |
181 | ** Note that degenerate cases are not only possible, but likely. | |
182 | ** For example, if the pair following b compares with opposite sense, | |
183 | ** then b == q < p == r == t. | |
184 | */ | |
185 | ||
186 | ||
957d8989 | 187 | static IV |
d4c19fe8 | 188 | dynprep(pTHX_ gptr *list1, gptr *list2, size_t nmemb, const SVCOMPARE_t cmp) |
84d4ea48 | 189 | { |
957d8989 | 190 | I32 sense; |
eb578fdb KW |
191 | gptr *b, *p, *q, *t, *p2; |
192 | gptr *last, *r; | |
957d8989 | 193 | IV runs = 0; |
84d4ea48 JH |
194 | |
195 | b = list1; | |
196 | last = PINDEX(b, nmemb); | |
197 | sense = (cmp(aTHX_ *b, *(b+1)) > 0); | |
198 | for (p2 = list2; b < last; ) { | |
199 | /* We just started, or just reversed sense. | |
200 | ** Set t at end of pairs with the prevailing sense. | |
201 | */ | |
202 | for (p = b+2, t = p; ++p < last; t = ++p) { | |
203 | if ((cmp(aTHX_ *t, *p) > 0) != sense) break; | |
204 | } | |
205 | q = b; | |
206 | /* Having laid out the playing field, look for long runs */ | |
207 | do { | |
208 | p = r = b + (2 * PTHRESH); | |
209 | if (r >= t) p = r = t; /* too short to care about */ | |
210 | else { | |
211 | while (((cmp(aTHX_ *(p-1), *p) > 0) == sense) && | |
47127b64 | 212 | ((p -= 2) > q)) {} |
84d4ea48 JH |
213 | if (p <= q) { |
214 | /* b through r is a (long) run. | |
215 | ** Extend it as far as possible. | |
216 | */ | |
217 | p = q = r; | |
218 | while (((p += 2) < t) && | |
219 | ((cmp(aTHX_ *(p-1), *p) > 0) == sense)) q = p; | |
220 | r = p = q + 2; /* no simple pairs, no after-run */ | |
221 | } | |
222 | } | |
223 | if (q > b) { /* run of greater than 2 at b */ | |
d4c19fe8 AL |
224 | gptr *savep = p; |
225 | ||
84d4ea48 JH |
226 | p = q += 2; |
227 | /* pick up singleton, if possible */ | |
228 | if ((p == t) && | |
229 | ((t + 1) == last) && | |
230 | ((cmp(aTHX_ *(p-1), *p) > 0) == sense)) | |
231 | savep = r = p = q = last; | |
957d8989 | 232 | p2 = NEXT(p2) = p2 + (p - b); ++runs; |
d4c19fe8 AL |
233 | if (sense) |
234 | while (b < --p) { | |
235 | const gptr c = *b; | |
236 | *b++ = *p; | |
237 | *p = c; | |
238 | } | |
84d4ea48 JH |
239 | p = savep; |
240 | } | |
241 | while (q < p) { /* simple pairs */ | |
957d8989 | 242 | p2 = NEXT(p2) = p2 + 2; ++runs; |
84d4ea48 | 243 | if (sense) { |
d4c19fe8 | 244 | const gptr c = *q++; |
84d4ea48 JH |
245 | *(q-1) = *q; |
246 | *q++ = c; | |
247 | } else q += 2; | |
248 | } | |
249 | if (((b = p) == t) && ((t+1) == last)) { | |
957d8989 | 250 | NEXT(p2) = p2 + 1; ++runs; |
84d4ea48 JH |
251 | b++; |
252 | } | |
253 | q = r; | |
254 | } while (b < t); | |
255 | sense = !sense; | |
256 | } | |
957d8989 | 257 | return runs; |
84d4ea48 JH |
258 | } |
259 | ||
260 | ||
3fe0b9a9 | 261 | /* The original merge sort, in use since 5.7, was as fast as, or faster than, |
957d8989 | 262 | * qsort on many platforms, but slower than qsort, conspicuously so, |
3fe0b9a9 | 263 | * on others. The most likely explanation was platform-specific |
957d8989 JL |
264 | * differences in cache sizes and relative speeds. |
265 | * | |
266 | * The quicksort divide-and-conquer algorithm guarantees that, as the | |
267 | * problem is subdivided into smaller and smaller parts, the parts | |
268 | * fit into smaller (and faster) caches. So it doesn't matter how | |
269 | * many levels of cache exist, quicksort will "find" them, and, | |
e62b3022 | 270 | * as long as smaller is faster, take advantage of them. |
957d8989 | 271 | * |
3fe0b9a9 | 272 | * By contrast, consider how the original mergesort algorithm worked. |
957d8989 JL |
273 | * Suppose we have five runs (each typically of length 2 after dynprep). |
274 | * | |
275 | * pass base aux | |
276 | * 0 1 2 3 4 5 | |
277 | * 1 12 34 5 | |
278 | * 2 1234 5 | |
279 | * 3 12345 | |
280 | * 4 12345 | |
281 | * | |
282 | * Adjacent pairs are merged in "grand sweeps" through the input. | |
283 | * This means, on pass 1, the records in runs 1 and 2 aren't revisited until | |
284 | * runs 3 and 4 are merged and the runs from run 5 have been copied. | |
285 | * The only cache that matters is one large enough to hold *all* the input. | |
286 | * On some platforms, this may be many times slower than smaller caches. | |
287 | * | |
288 | * The following pseudo-code uses the same basic merge algorithm, | |
289 | * but in a divide-and-conquer way. | |
290 | * | |
291 | * # merge $runs runs at offset $offset of list $list1 into $list2. | |
292 | * # all unmerged runs ($runs == 1) originate in list $base. | |
293 | * sub mgsort2 { | |
294 | * my ($offset, $runs, $base, $list1, $list2) = @_; | |
295 | * | |
296 | * if ($runs == 1) { | |
297 | * if ($list1 is $base) copy run to $list2 | |
298 | * return offset of end of list (or copy) | |
299 | * } else { | |
300 | * $off2 = mgsort2($offset, $runs-($runs/2), $base, $list2, $list1) | |
301 | * mgsort2($off2, $runs/2, $base, $list2, $list1) | |
302 | * merge the adjacent runs at $offset of $list1 into $list2 | |
303 | * return the offset of the end of the merged runs | |
304 | * } | |
305 | * } | |
306 | * mgsort2(0, $runs, $base, $aux, $base); | |
307 | * | |
308 | * For our 5 runs, the tree of calls looks like | |
309 | * | |
310 | * 5 | |
311 | * 3 2 | |
312 | * 2 1 1 1 | |
313 | * 1 1 | |
314 | * | |
315 | * 1 2 3 4 5 | |
316 | * | |
317 | * and the corresponding activity looks like | |
318 | * | |
319 | * copy runs 1 and 2 from base to aux | |
320 | * merge runs 1 and 2 from aux to base | |
321 | * (run 3 is where it belongs, no copy needed) | |
322 | * merge runs 12 and 3 from base to aux | |
323 | * (runs 4 and 5 are where they belong, no copy needed) | |
324 | * merge runs 4 and 5 from base to aux | |
325 | * merge runs 123 and 45 from aux to base | |
326 | * | |
327 | * Note that we merge runs 1 and 2 immediately after copying them, | |
328 | * while they are still likely to be in fast cache. Similarly, | |
329 | * run 3 is merged with run 12 while it still may be lingering in cache. | |
330 | * This implementation should therefore enjoy much of the cache-friendly | |
331 | * behavior that quicksort does. In addition, it does less copying | |
332 | * than the original mergesort implementation (only runs 1 and 2 are copied) | |
333 | * and the "balancing" of merges is better (merged runs comprise more nearly | |
334 | * equal numbers of original runs). | |
335 | * | |
336 | * The actual cache-friendly implementation will use a pseudo-stack | |
337 | * to avoid recursion, and will unroll processing of runs of length 2, | |
338 | * but it is otherwise similar to the recursive implementation. | |
957d8989 JL |
339 | */ |
340 | ||
341 | typedef struct { | |
342 | IV offset; /* offset of 1st of 2 runs at this level */ | |
343 | IV runs; /* how many runs must be combined into 1 */ | |
344 | } off_runs; /* pseudo-stack element */ | |
345 | ||
6c3fb703 NC |
346 | |
347 | static I32 | |
31e9e0a3 | 348 | cmp_desc(pTHX_ gptr const a, gptr const b) |
6c3fb703 NC |
349 | { |
350 | return -PL_sort_RealCmp(aTHX_ a, b); | |
351 | } | |
352 | ||
e2091bb6 Z |
353 | /* |
354 | =for apidoc sortsv_flags | |
355 | ||
356 | In-place sort an array of SV pointers with the given comparison routine, | |
357 | with various SORTf_* flag options. | |
358 | ||
359 | =cut | |
360 | */ | |
361 | void | |
362 | Perl_sortsv_flags(pTHX_ gptr *base, size_t nmemb, SVCOMPARE_t cmp, U32 flags) | |
957d8989 | 363 | { |
551405c4 | 364 | IV i, run, offset; |
957d8989 | 365 | I32 sense, level; |
eb578fdb | 366 | gptr *f1, *f2, *t, *b, *p; |
957d8989 | 367 | int iwhich; |
551405c4 | 368 | gptr *aux; |
957d8989 JL |
369 | gptr *p1; |
370 | gptr small[SMALLSORT]; | |
371 | gptr *which[3]; | |
372 | off_runs stack[60], *stackp; | |
d4c19fe8 | 373 | SVCOMPARE_t savecmp = NULL; |
957d8989 | 374 | |
e2091bb6 | 375 | PERL_ARGS_ASSERT_SORTSV_FLAGS; |
957d8989 | 376 | if (nmemb <= 1) return; /* sorted trivially */ |
6c3fb703 | 377 | |
f4f44d65 | 378 | if ((flags & SORTf_DESC) != 0) { |
6c3fb703 NC |
379 | savecmp = PL_sort_RealCmp; /* Save current comparison routine, if any */ |
380 | PL_sort_RealCmp = cmp; /* Put comparison routine where cmp_desc can find it */ | |
381 | cmp = cmp_desc; | |
382 | } | |
383 | ||
957d8989 | 384 | if (nmemb <= SMALLSORT) aux = small; /* use stack for aux array */ |
486ec47a | 385 | else { Newx(aux,nmemb,gptr); } /* allocate auxiliary array */ |
957d8989 JL |
386 | level = 0; |
387 | stackp = stack; | |
388 | stackp->runs = dynprep(aTHX_ base, aux, nmemb, cmp); | |
389 | stackp->offset = offset = 0; | |
390 | which[0] = which[2] = base; | |
391 | which[1] = aux; | |
392 | for (;;) { | |
393 | /* On levels where both runs have be constructed (stackp->runs == 0), | |
394 | * merge them, and note the offset of their end, in case the offset | |
395 | * is needed at the next level up. Hop up a level, and, | |
396 | * as long as stackp->runs is 0, keep merging. | |
397 | */ | |
551405c4 AL |
398 | IV runs = stackp->runs; |
399 | if (runs == 0) { | |
400 | gptr *list1, *list2; | |
957d8989 JL |
401 | iwhich = level & 1; |
402 | list1 = which[iwhich]; /* area where runs are now */ | |
403 | list2 = which[++iwhich]; /* area for merged runs */ | |
404 | do { | |
eb578fdb | 405 | gptr *l1, *l2, *tp2; |
957d8989 JL |
406 | offset = stackp->offset; |
407 | f1 = p1 = list1 + offset; /* start of first run */ | |
408 | p = tp2 = list2 + offset; /* where merged run will go */ | |
409 | t = NEXT(p); /* where first run ends */ | |
410 | f2 = l1 = POTHER(t, list2, list1); /* ... on the other side */ | |
411 | t = NEXT(t); /* where second runs ends */ | |
412 | l2 = POTHER(t, list2, list1); /* ... on the other side */ | |
413 | offset = PNELEM(list2, t); | |
414 | while (f1 < l1 && f2 < l2) { | |
415 | /* If head 1 is larger than head 2, find ALL the elements | |
416 | ** in list 2 strictly less than head1, write them all, | |
417 | ** then head 1. Then compare the new heads, and repeat, | |
418 | ** until one or both lists are exhausted. | |
419 | ** | |
420 | ** In all comparisons (after establishing | |
421 | ** which head to merge) the item to merge | |
422 | ** (at pointer q) is the first operand of | |
423 | ** the comparison. When we want to know | |
a0288114 | 424 | ** if "q is strictly less than the other", |
957d8989 JL |
425 | ** we can't just do |
426 | ** cmp(q, other) < 0 | |
427 | ** because stability demands that we treat equality | |
428 | ** as high when q comes from l2, and as low when | |
429 | ** q was from l1. So we ask the question by doing | |
430 | ** cmp(q, other) <= sense | |
431 | ** and make sense == 0 when equality should look low, | |
432 | ** and -1 when equality should look high. | |
433 | */ | |
434 | ||
eb578fdb | 435 | gptr *q; |
957d8989 JL |
436 | if (cmp(aTHX_ *f1, *f2) <= 0) { |
437 | q = f2; b = f1; t = l1; | |
438 | sense = -1; | |
439 | } else { | |
440 | q = f1; b = f2; t = l2; | |
441 | sense = 0; | |
442 | } | |
443 | ||
444 | ||
445 | /* ramp up | |
446 | ** | |
447 | ** Leave t at something strictly | |
448 | ** greater than q (or at the end of the list), | |
449 | ** and b at something strictly less than q. | |
450 | */ | |
451 | for (i = 1, run = 0 ;;) { | |
452 | if ((p = PINDEX(b, i)) >= t) { | |
453 | /* off the end */ | |
454 | if (((p = PINDEX(t, -1)) > b) && | |
455 | (cmp(aTHX_ *q, *p) <= sense)) | |
456 | t = p; | |
457 | else b = p; | |
458 | break; | |
459 | } else if (cmp(aTHX_ *q, *p) <= sense) { | |
460 | t = p; | |
461 | break; | |
462 | } else b = p; | |
463 | if (++run >= RTHRESH) i += i; | |
464 | } | |
465 | ||
466 | ||
467 | /* q is known to follow b and must be inserted before t. | |
468 | ** Increment b, so the range of possibilities is [b,t). | |
469 | ** Round binary split down, to favor early appearance. | |
470 | ** Adjust b and t until q belongs just before t. | |
471 | */ | |
472 | ||
473 | b++; | |
474 | while (b < t) { | |
475 | p = PINDEX(b, (PNELEM(b, t) - 1) / 2); | |
476 | if (cmp(aTHX_ *q, *p) <= sense) { | |
477 | t = p; | |
478 | } else b = p + 1; | |
479 | } | |
480 | ||
481 | ||
482 | /* Copy all the strictly low elements */ | |
483 | ||
484 | if (q == f1) { | |
485 | FROMTOUPTO(f2, tp2, t); | |
486 | *tp2++ = *f1++; | |
487 | } else { | |
488 | FROMTOUPTO(f1, tp2, t); | |
489 | *tp2++ = *f2++; | |
490 | } | |
491 | } | |
492 | ||
493 | ||
494 | /* Run out remaining list */ | |
495 | if (f1 == l1) { | |
496 | if (f2 < l2) FROMTOUPTO(f2, tp2, l2); | |
497 | } else FROMTOUPTO(f1, tp2, l1); | |
498 | p1 = NEXT(p1) = POTHER(tp2, list2, list1); | |
499 | ||
500 | if (--level == 0) goto done; | |
501 | --stackp; | |
502 | t = list1; list1 = list2; list2 = t; /* swap lists */ | |
503 | } while ((runs = stackp->runs) == 0); | |
504 | } | |
505 | ||
506 | ||
507 | stackp->runs = 0; /* current run will finish level */ | |
508 | /* While there are more than 2 runs remaining, | |
509 | * turn them into exactly 2 runs (at the "other" level), | |
510 | * each made up of approximately half the runs. | |
511 | * Stack the second half for later processing, | |
512 | * and set about producing the first half now. | |
513 | */ | |
514 | while (runs > 2) { | |
515 | ++level; | |
516 | ++stackp; | |
517 | stackp->offset = offset; | |
518 | runs -= stackp->runs = runs / 2; | |
519 | } | |
520 | /* We must construct a single run from 1 or 2 runs. | |
521 | * All the original runs are in which[0] == base. | |
522 | * The run we construct must end up in which[level&1]. | |
523 | */ | |
524 | iwhich = level & 1; | |
525 | if (runs == 1) { | |
526 | /* Constructing a single run from a single run. | |
527 | * If it's where it belongs already, there's nothing to do. | |
528 | * Otherwise, copy it to where it belongs. | |
529 | * A run of 1 is either a singleton at level 0, | |
530 | * or the second half of a split 3. In neither event | |
531 | * is it necessary to set offset. It will be set by the merge | |
532 | * that immediately follows. | |
533 | */ | |
534 | if (iwhich) { /* Belongs in aux, currently in base */ | |
535 | f1 = b = PINDEX(base, offset); /* where list starts */ | |
536 | f2 = PINDEX(aux, offset); /* where list goes */ | |
537 | t = NEXT(f2); /* where list will end */ | |
538 | offset = PNELEM(aux, t); /* offset thereof */ | |
539 | t = PINDEX(base, offset); /* where it currently ends */ | |
540 | FROMTOUPTO(f1, f2, t); /* copy */ | |
541 | NEXT(b) = t; /* set up parallel pointer */ | |
542 | } else if (level == 0) goto done; /* single run at level 0 */ | |
543 | } else { | |
544 | /* Constructing a single run from two runs. | |
545 | * The merge code at the top will do that. | |
546 | * We need only make sure the two runs are in the "other" array, | |
547 | * so they'll end up in the correct array after the merge. | |
548 | */ | |
549 | ++level; | |
550 | ++stackp; | |
551 | stackp->offset = offset; | |
552 | stackp->runs = 0; /* take care of both runs, trigger merge */ | |
553 | if (!iwhich) { /* Merged runs belong in aux, copy 1st */ | |
554 | f1 = b = PINDEX(base, offset); /* where first run starts */ | |
555 | f2 = PINDEX(aux, offset); /* where it will be copied */ | |
556 | t = NEXT(f2); /* where first run will end */ | |
557 | offset = PNELEM(aux, t); /* offset thereof */ | |
558 | p = PINDEX(base, offset); /* end of first run */ | |
559 | t = NEXT(t); /* where second run will end */ | |
560 | t = PINDEX(base, PNELEM(aux, t)); /* where it now ends */ | |
561 | FROMTOUPTO(f1, f2, t); /* copy both runs */ | |
486ec47a | 562 | NEXT(b) = p; /* paralleled pointer for 1st */ |
957d8989 JL |
563 | NEXT(p) = t; /* ... and for second */ |
564 | } | |
565 | } | |
566 | } | |
7b52d656 | 567 | done: |
957d8989 | 568 | if (aux != small) Safefree(aux); /* free iff allocated */ |
0e1d050c | 569 | if (savecmp != NULL) { |
6c3fb703 NC |
570 | PL_sort_RealCmp = savecmp; /* Restore current comparison routine, if any */ |
571 | } | |
957d8989 JL |
572 | return; |
573 | } | |
574 | ||
84d4ea48 JH |
575 | /* |
576 | * The quicksort implementation was derived from source code contributed | |
577 | * by Tom Horsley. | |
578 | * | |
579 | * NOTE: this code was derived from Tom Horsley's qsort replacement | |
580 | * and should not be confused with the original code. | |
581 | */ | |
582 | ||
583 | /* Copyright (C) Tom Horsley, 1997. All rights reserved. | |
584 | ||
585 | Permission granted to distribute under the same terms as perl which are | |
586 | (briefly): | |
587 | ||
588 | This program is free software; you can redistribute it and/or modify | |
589 | it under the terms of either: | |
590 | ||
591 | a) the GNU General Public License as published by the Free | |
592 | Software Foundation; either version 1, or (at your option) any | |
593 | later version, or | |
594 | ||
595 | b) the "Artistic License" which comes with this Kit. | |
596 | ||
597 | Details on the perl license can be found in the perl source code which | |
598 | may be located via the www.perl.com web page. | |
599 | ||
600 | This is the most wonderfulest possible qsort I can come up with (and | |
601 | still be mostly portable) My (limited) tests indicate it consistently | |
602 | does about 20% fewer calls to compare than does the qsort in the Visual | |
603 | C++ library, other vendors may vary. | |
604 | ||
605 | Some of the ideas in here can be found in "Algorithms" by Sedgewick, | |
606 | others I invented myself (or more likely re-invented since they seemed | |
607 | pretty obvious once I watched the algorithm operate for a while). | |
608 | ||
609 | Most of this code was written while watching the Marlins sweep the Giants | |
610 | in the 1997 National League Playoffs - no Braves fans allowed to use this | |
611 | code (just kidding :-). | |
612 | ||
613 | I realize that if I wanted to be true to the perl tradition, the only | |
614 | comment in this file would be something like: | |
615 | ||
616 | ...they shuffled back towards the rear of the line. 'No, not at the | |
617 | rear!' the slave-driver shouted. 'Three files up. And stay there... | |
618 | ||
619 | However, I really needed to violate that tradition just so I could keep | |
620 | track of what happens myself, not to mention some poor fool trying to | |
621 | understand this years from now :-). | |
622 | */ | |
623 | ||
624 | /* ********************************************************** Configuration */ | |
625 | ||
626 | #ifndef QSORT_ORDER_GUESS | |
627 | #define QSORT_ORDER_GUESS 2 /* Select doubling version of the netBSD trick */ | |
628 | #endif | |
629 | ||
630 | /* QSORT_MAX_STACK is the largest number of partitions that can be stacked up for | |
631 | future processing - a good max upper bound is log base 2 of memory size | |
632 | (32 on 32 bit machines, 64 on 64 bit machines, etc). In reality can | |
633 | safely be smaller than that since the program is taking up some space and | |
634 | most operating systems only let you grab some subset of contiguous | |
635 | memory (not to mention that you are normally sorting data larger than | |
636 | 1 byte element size :-). | |
637 | */ | |
638 | #ifndef QSORT_MAX_STACK | |
639 | #define QSORT_MAX_STACK 32 | |
640 | #endif | |
641 | ||
642 | /* QSORT_BREAK_EVEN is the size of the largest partition we should insertion sort. | |
643 | Anything bigger and we use qsort. If you make this too small, the qsort | |
644 | will probably break (or become less efficient), because it doesn't expect | |
645 | the middle element of a partition to be the same as the right or left - | |
646 | you have been warned). | |
647 | */ | |
648 | #ifndef QSORT_BREAK_EVEN | |
649 | #define QSORT_BREAK_EVEN 6 | |
650 | #endif | |
651 | ||
4eb872f6 JL |
652 | /* QSORT_PLAY_SAFE is the size of the largest partition we're willing |
653 | to go quadratic on. We innoculate larger partitions against | |
654 | quadratic behavior by shuffling them before sorting. This is not | |
655 | an absolute guarantee of non-quadratic behavior, but it would take | |
656 | staggeringly bad luck to pick extreme elements as the pivot | |
657 | from randomized data. | |
658 | */ | |
659 | #ifndef QSORT_PLAY_SAFE | |
660 | #define QSORT_PLAY_SAFE 255 | |
661 | #endif | |
662 | ||
84d4ea48 JH |
663 | /* ************************************************************* Data Types */ |
664 | ||
665 | /* hold left and right index values of a partition waiting to be sorted (the | |
666 | partition includes both left and right - right is NOT one past the end or | |
667 | anything like that). | |
668 | */ | |
669 | struct partition_stack_entry { | |
670 | int left; | |
671 | int right; | |
672 | #ifdef QSORT_ORDER_GUESS | |
673 | int qsort_break_even; | |
674 | #endif | |
675 | }; | |
676 | ||
677 | /* ******************************************************* Shorthand Macros */ | |
678 | ||
679 | /* Note that these macros will be used from inside the qsort function where | |
680 | we happen to know that the variable 'elt_size' contains the size of an | |
681 | array element and the variable 'temp' points to enough space to hold a | |
682 | temp element and the variable 'array' points to the array being sorted | |
683 | and 'compare' is the pointer to the compare routine. | |
684 | ||
685 | Also note that there are very many highly architecture specific ways | |
686 | these might be sped up, but this is simply the most generally portable | |
687 | code I could think of. | |
688 | */ | |
689 | ||
690 | /* Return < 0 == 0 or > 0 as the value of elt1 is < elt2, == elt2, > elt2 | |
691 | */ | |
692 | #define qsort_cmp(elt1, elt2) \ | |
693 | ((*compare)(aTHX_ array[elt1], array[elt2])) | |
694 | ||
695 | #ifdef QSORT_ORDER_GUESS | |
696 | #define QSORT_NOTICE_SWAP swapped++; | |
697 | #else | |
698 | #define QSORT_NOTICE_SWAP | |
699 | #endif | |
700 | ||
701 | /* swaps contents of array elements elt1, elt2. | |
702 | */ | |
703 | #define qsort_swap(elt1, elt2) \ | |
704 | STMT_START { \ | |
705 | QSORT_NOTICE_SWAP \ | |
706 | temp = array[elt1]; \ | |
707 | array[elt1] = array[elt2]; \ | |
708 | array[elt2] = temp; \ | |
709 | } STMT_END | |
710 | ||
711 | /* rotate contents of elt1, elt2, elt3 such that elt1 gets elt2, elt2 gets | |
712 | elt3 and elt3 gets elt1. | |
713 | */ | |
714 | #define qsort_rotate(elt1, elt2, elt3) \ | |
715 | STMT_START { \ | |
716 | QSORT_NOTICE_SWAP \ | |
717 | temp = array[elt1]; \ | |
718 | array[elt1] = array[elt2]; \ | |
719 | array[elt2] = array[elt3]; \ | |
720 | array[elt3] = temp; \ | |
721 | } STMT_END | |
722 | ||
723 | /* ************************************************************ Debug stuff */ | |
724 | ||
725 | #ifdef QSORT_DEBUG | |
726 | ||
727 | static void | |
728 | break_here() | |
729 | { | |
730 | return; /* good place to set a breakpoint */ | |
731 | } | |
732 | ||
733 | #define qsort_assert(t) (void)( (t) || (break_here(), 0) ) | |
734 | ||
735 | static void | |
736 | doqsort_all_asserts( | |
737 | void * array, | |
738 | size_t num_elts, | |
739 | size_t elt_size, | |
740 | int (*compare)(const void * elt1, const void * elt2), | |
741 | int pc_left, int pc_right, int u_left, int u_right) | |
742 | { | |
743 | int i; | |
744 | ||
745 | qsort_assert(pc_left <= pc_right); | |
746 | qsort_assert(u_right < pc_left); | |
747 | qsort_assert(pc_right < u_left); | |
748 | for (i = u_right + 1; i < pc_left; ++i) { | |
749 | qsort_assert(qsort_cmp(i, pc_left) < 0); | |
750 | } | |
751 | for (i = pc_left; i < pc_right; ++i) { | |
752 | qsort_assert(qsort_cmp(i, pc_right) == 0); | |
753 | } | |
754 | for (i = pc_right + 1; i < u_left; ++i) { | |
755 | qsort_assert(qsort_cmp(pc_right, i) < 0); | |
756 | } | |
757 | } | |
758 | ||
759 | #define qsort_all_asserts(PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT) \ | |
760 | doqsort_all_asserts(array, num_elts, elt_size, compare, \ | |
761 | PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT) | |
762 | ||
763 | #else | |
764 | ||
765 | #define qsort_assert(t) ((void)0) | |
766 | ||
767 | #define qsort_all_asserts(PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT) ((void)0) | |
768 | ||
769 | #endif | |
770 | ||
4eb872f6 | 771 | /* |
ccfc67b7 JH |
772 | =head1 Array Manipulation Functions |
773 | ||
84d4ea48 JH |
774 | =for apidoc sortsv |
775 | ||
8f5d5a51 | 776 | In-place sort an array of SV pointers with the given comparison routine. |
84d4ea48 | 777 | |
796b6530 | 778 | Currently this always uses mergesort. See C<L</sortsv_flags>> for a more |
7b9ef140 | 779 | flexible routine. |
78210658 | 780 | |
84d4ea48 JH |
781 | =cut |
782 | */ | |
4eb872f6 | 783 | |
84d4ea48 JH |
784 | void |
785 | Perl_sortsv(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp) | |
786 | { | |
7918f24d NC |
787 | PERL_ARGS_ASSERT_SORTSV; |
788 | ||
7b9ef140 | 789 | sortsv_flags(array, nmemb, cmp, 0); |
6c3fb703 NC |
790 | } |
791 | ||
4d562308 SF |
792 | #define SvNSIOK(sv) ((SvFLAGS(sv) & SVf_NOK) || ((SvFLAGS(sv) & (SVf_IOK|SVf_IVisUV)) == SVf_IOK)) |
793 | #define SvSIOK(sv) ((SvFLAGS(sv) & (SVf_IOK|SVf_IVisUV)) == SVf_IOK) | |
794 | #define SvNSIV(sv) ( SvNOK(sv) ? SvNVX(sv) : ( SvSIOK(sv) ? SvIVX(sv) : sv_2nv(sv) ) ) | |
795 | ||
84d4ea48 JH |
796 | PP(pp_sort) |
797 | { | |
20b7effb | 798 | dSP; dMARK; dORIGMARK; |
eb578fdb | 799 | SV **p1 = ORIGMARK+1, **p2; |
c70927a6 | 800 | SSize_t max, i; |
7d49f689 | 801 | AV* av = NULL; |
84d4ea48 | 802 | GV *gv; |
cbbf8932 | 803 | CV *cv = NULL; |
1c23e2bd | 804 | U8 gimme = GIMME_V; |
0bcc34c2 | 805 | OP* const nextop = PL_op->op_next; |
84d4ea48 JH |
806 | I32 overloading = 0; |
807 | bool hasargs = FALSE; | |
2b66f6d3 | 808 | bool copytmps; |
84d4ea48 | 809 | I32 is_xsub = 0; |
901017d6 AL |
810 | const U8 priv = PL_op->op_private; |
811 | const U8 flags = PL_op->op_flags; | |
7b9ef140 RH |
812 | U32 sort_flags = 0; |
813 | void (*sortsvp)(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp, U32 flags) | |
814 | = Perl_sortsv_flags; | |
4d562308 | 815 | I32 all_SIVs = 1; |
84d4ea48 | 816 | |
7b9ef140 RH |
817 | if ((priv & OPpSORT_DESCEND) != 0) |
818 | sort_flags |= SORTf_DESC; | |
7b9ef140 RH |
819 | if ((priv & OPpSORT_STABLE) != 0) |
820 | sort_flags |= SORTf_STABLE; | |
afe59f35 FC |
821 | if ((priv & OPpSORT_UNSTABLE) != 0) |
822 | sort_flags |= SORTf_UNSTABLE; | |
7b9ef140 | 823 | |
84d4ea48 JH |
824 | if (gimme != G_ARRAY) { |
825 | SP = MARK; | |
b59aed67 | 826 | EXTEND(SP,1); |
84d4ea48 JH |
827 | RETPUSHUNDEF; |
828 | } | |
829 | ||
830 | ENTER; | |
831 | SAVEVPTR(PL_sortcop); | |
471178c0 NC |
832 | if (flags & OPf_STACKED) { |
833 | if (flags & OPf_SPECIAL) { | |
e6dae479 | 834 | OP *nullop = OpSIBLING(cLISTOP->op_first); /* pass pushmark */ |
932bca29 DM |
835 | assert(nullop->op_type == OP_NULL); |
836 | PL_sortcop = nullop->op_next; | |
84d4ea48 JH |
837 | } |
838 | else { | |
f7bc00ea | 839 | GV *autogv = NULL; |
5a34f1cd | 840 | HV *stash; |
f7bc00ea FC |
841 | cv = sv_2cv(*++MARK, &stash, &gv, GV_ADD); |
842 | check_cv: | |
84d4ea48 | 843 | if (cv && SvPOK(cv)) { |
ad64d0ec | 844 | const char * const proto = SvPV_nolen_const(MUTABLE_SV(cv)); |
84d4ea48 JH |
845 | if (proto && strEQ(proto, "$$")) { |
846 | hasargs = TRUE; | |
847 | } | |
848 | } | |
2fc49ef1 FC |
849 | if (cv && CvISXSUB(cv) && CvXSUB(cv)) { |
850 | is_xsub = 1; | |
851 | } | |
852 | else if (!(cv && CvROOT(cv))) { | |
853 | if (gv) { | |
f7bc00ea FC |
854 | goto autoload; |
855 | } | |
856 | else if (!CvANON(cv) && (gv = CvGV(cv))) { | |
857 | if (cv != GvCV(gv)) cv = GvCV(gv); | |
858 | autoload: | |
859 | if (!autogv && ( | |
860 | autogv = gv_autoload_pvn( | |
861 | GvSTASH(gv), GvNAME(gv), GvNAMELEN(gv), | |
862 | GvNAMEUTF8(gv) ? SVf_UTF8 : 0 | |
863 | ) | |
864 | )) { | |
865 | cv = GvCVu(autogv); | |
866 | goto check_cv; | |
867 | } | |
868 | else { | |
84d4ea48 | 869 | SV *tmpstr = sv_newmortal(); |
bd61b366 | 870 | gv_efullname3(tmpstr, gv, NULL); |
147e3846 | 871 | DIE(aTHX_ "Undefined sort subroutine \"%" SVf "\" called", |
be2597df | 872 | SVfARG(tmpstr)); |
f7bc00ea | 873 | } |
84d4ea48 JH |
874 | } |
875 | else { | |
876 | DIE(aTHX_ "Undefined subroutine in sort"); | |
877 | } | |
878 | } | |
879 | ||
880 | if (is_xsub) | |
881 | PL_sortcop = (OP*)cv; | |
9850bf21 | 882 | else |
84d4ea48 | 883 | PL_sortcop = CvSTART(cv); |
84d4ea48 JH |
884 | } |
885 | } | |
886 | else { | |
5f66b61c | 887 | PL_sortcop = NULL; |
84d4ea48 JH |
888 | } |
889 | ||
84721d61 DM |
890 | /* optimiser converts "@a = sort @a" to "sort \@a". In this case, |
891 | * push (@a) onto stack, then assign result back to @a at the end of | |
892 | * this function */ | |
0723351e | 893 | if (priv & OPpSORT_INPLACE) { |
fe1bc4cf DM |
894 | assert( MARK+1 == SP && *SP && SvTYPE(*SP) == SVt_PVAV); |
895 | (void)POPMARK; /* remove mark associated with ex-OP_AASSIGN */ | |
502c6561 | 896 | av = MUTABLE_AV((*SP)); |
84721d61 DM |
897 | if (SvREADONLY(av)) |
898 | Perl_croak_no_modify(); | |
fe1bc4cf | 899 | max = AvFILL(av) + 1; |
84721d61 | 900 | MEXTEND(SP, max); |
fe1bc4cf | 901 | if (SvMAGICAL(av)) { |
fe2774ed | 902 | for (i=0; i < max; i++) { |
fe1bc4cf | 903 | SV **svp = av_fetch(av, i, FALSE); |
a0714e2c | 904 | *SP++ = (svp) ? *svp : NULL; |
fe1bc4cf DM |
905 | } |
906 | } | |
84721d61 DM |
907 | else { |
908 | SV **svp = AvARRAY(av); | |
909 | assert(svp || max == 0); | |
910 | for (i = 0; i < max; i++) | |
911 | *SP++ = *svp++; | |
fe1bc4cf | 912 | } |
84721d61 DM |
913 | SP--; |
914 | p1 = p2 = SP - (max-1); | |
fe1bc4cf DM |
915 | } |
916 | else { | |
917 | p2 = MARK+1; | |
918 | max = SP - MARK; | |
919 | } | |
920 | ||
83a44efe SF |
921 | /* shuffle stack down, removing optional initial cv (p1!=p2), plus |
922 | * any nulls; also stringify or converting to integer or number as | |
923 | * required any args */ | |
ff859a7f | 924 | copytmps = cBOOL(PL_sortcop); |
fe1bc4cf DM |
925 | for (i=max; i > 0 ; i--) { |
926 | if ((*p1 = *p2++)) { /* Weed out nulls. */ | |
60779a30 | 927 | if (copytmps && SvPADTMP(*p1)) { |
2b66f6d3 | 928 | *p1 = sv_mortalcopy(*p1); |
60779a30 | 929 | } |
fe1bc4cf | 930 | SvTEMP_off(*p1); |
83a44efe SF |
931 | if (!PL_sortcop) { |
932 | if (priv & OPpSORT_NUMERIC) { | |
933 | if (priv & OPpSORT_INTEGER) { | |
bdbefedf DM |
934 | if (!SvIOK(*p1)) |
935 | (void)sv_2iv_flags(*p1, SV_GMAGIC|SV_SKIP_OVERLOAD); | |
83a44efe SF |
936 | } |
937 | else { | |
bdbefedf DM |
938 | if (!SvNSIOK(*p1)) |
939 | (void)sv_2nv_flags(*p1, SV_GMAGIC|SV_SKIP_OVERLOAD); | |
4d562308 SF |
940 | if (all_SIVs && !SvSIOK(*p1)) |
941 | all_SIVs = 0; | |
83a44efe SF |
942 | } |
943 | } | |
944 | else { | |
bdbefedf DM |
945 | if (!SvPOK(*p1)) |
946 | (void)sv_2pv_flags(*p1, 0, | |
947 | SV_GMAGIC|SV_CONST_RETURN|SV_SKIP_OVERLOAD); | |
83a44efe | 948 | } |
bdbefedf DM |
949 | if (SvAMAGIC(*p1)) |
950 | overloading = 1; | |
84d4ea48 | 951 | } |
fe1bc4cf | 952 | p1++; |
84d4ea48 | 953 | } |
fe1bc4cf DM |
954 | else |
955 | max--; | |
84d4ea48 | 956 | } |
fe1bc4cf | 957 | if (max > 1) { |
471178c0 | 958 | SV **start; |
fe1bc4cf | 959 | if (PL_sortcop) { |
84d4ea48 | 960 | PERL_CONTEXT *cx; |
901017d6 | 961 | const bool oldcatch = CATCH_GET; |
8ae997c5 | 962 | I32 old_savestack_ix = PL_savestack_ix; |
84d4ea48 | 963 | |
84d4ea48 JH |
964 | SAVEOP(); |
965 | ||
966 | CATCH_SET(TRUE); | |
967 | PUSHSTACKi(PERLSI_SORT); | |
968 | if (!hasargs && !is_xsub) { | |
8465ba45 FC |
969 | SAVEGENERICSV(PL_firstgv); |
970 | SAVEGENERICSV(PL_secondgv); | |
971 | PL_firstgv = MUTABLE_GV(SvREFCNT_inc( | |
972 | gv_fetchpvs("a", GV_ADD|GV_NOTQUAL, SVt_PV) | |
973 | )); | |
974 | PL_secondgv = MUTABLE_GV(SvREFCNT_inc( | |
975 | gv_fetchpvs("b", GV_ADD|GV_NOTQUAL, SVt_PV) | |
976 | )); | |
dc9ef998 TC |
977 | /* make sure the GP isn't removed out from under us for |
978 | * the SAVESPTR() */ | |
979 | save_gp(PL_firstgv, 0); | |
980 | save_gp(PL_secondgv, 0); | |
981 | /* we don't want modifications localized */ | |
982 | GvINTRO_off(PL_firstgv); | |
983 | GvINTRO_off(PL_secondgv); | |
16ada235 Z |
984 | SAVEGENERICSV(GvSV(PL_firstgv)); |
985 | SvREFCNT_inc(GvSV(PL_firstgv)); | |
986 | SAVEGENERICSV(GvSV(PL_secondgv)); | |
987 | SvREFCNT_inc(GvSV(PL_secondgv)); | |
84d4ea48 JH |
988 | } |
989 | ||
33411212 | 990 | gimme = G_SCALAR; |
ed8ff0f3 | 991 | cx = cx_pushblock(CXt_NULL, gimme, PL_stack_base, old_savestack_ix); |
471178c0 | 992 | if (!(flags & OPf_SPECIAL)) { |
79646418 | 993 | cx->cx_type = CXt_SUB|CXp_MULTICALL; |
a73d8813 | 994 | cx_pushsub(cx, cv, NULL, hasargs); |
9850bf21 | 995 | if (!is_xsub) { |
b70d5558 | 996 | PADLIST * const padlist = CvPADLIST(cv); |
9850bf21 | 997 | |
d2af2719 | 998 | if (++CvDEPTH(cv) >= 2) |
9850bf21 | 999 | pad_push(padlist, CvDEPTH(cv)); |
9850bf21 | 1000 | PAD_SET_CUR_NOSAVE(padlist, CvDEPTH(cv)); |
84d4ea48 | 1001 | |
9850bf21 RH |
1002 | if (hasargs) { |
1003 | /* This is mostly copied from pp_entersub */ | |
502c6561 | 1004 | AV * const av = MUTABLE_AV(PAD_SVl(0)); |
84d4ea48 | 1005 | |
9850bf21 | 1006 | cx->blk_sub.savearray = GvAV(PL_defgv); |
502c6561 | 1007 | GvAV(PL_defgv) = MUTABLE_AV(SvREFCNT_inc_simple(av)); |
9850bf21 RH |
1008 | } |
1009 | ||
1010 | } | |
84d4ea48 | 1011 | } |
486430a5 | 1012 | |
471178c0 NC |
1013 | start = p1 - max; |
1014 | sortsvp(aTHX_ start, max, | |
7b9ef140 RH |
1015 | (is_xsub ? S_sortcv_xsub : hasargs ? S_sortcv_stacked : S_sortcv), |
1016 | sort_flags); | |
84d4ea48 | 1017 | |
4df352a8 | 1018 | /* Reset cx, in case the context stack has been reallocated. */ |
4ebe6e95 | 1019 | cx = CX_CUR(); |
4df352a8 DM |
1020 | |
1021 | PL_stack_sp = PL_stack_base + cx->blk_oldsp; | |
1022 | ||
2f450c1b | 1023 | CX_LEAVE_SCOPE(cx); |
9850bf21 | 1024 | if (!(flags & OPf_SPECIAL)) { |
4df352a8 | 1025 | assert(CxTYPE(cx) == CXt_SUB); |
a73d8813 | 1026 | cx_popsub(cx); |
9850bf21 | 1027 | } |
2f450c1b | 1028 | else |
4df352a8 | 1029 | assert(CxTYPE(cx) == CXt_NULL); |
2f450c1b | 1030 | /* there isn't a POPNULL ! */ |
1dfbe6b4 | 1031 | |
ed8ff0f3 | 1032 | cx_popblock(cx); |
5da525e9 | 1033 | CX_POP(cx); |
84d4ea48 JH |
1034 | POPSTACK; |
1035 | CATCH_SET(oldcatch); | |
1036 | } | |
fe1bc4cf | 1037 | else { |
84d4ea48 | 1038 | MEXTEND(SP, 20); /* Can't afford stack realloc on signal. */ |
84721d61 | 1039 | start = ORIGMARK+1; |
471178c0 | 1040 | sortsvp(aTHX_ start, max, |
0723351e | 1041 | (priv & OPpSORT_NUMERIC) |
4d562308 | 1042 | ? ( ( ( priv & OPpSORT_INTEGER) || all_SIVs) |
f0f5dc9d AL |
1043 | ? ( overloading ? S_amagic_i_ncmp : S_sv_i_ncmp) |
1044 | : ( overloading ? S_amagic_ncmp : S_sv_ncmp ) ) | |
130c5df3 KW |
1045 | : ( |
1046 | #ifdef USE_LOCALE_COLLATE | |
1047 | IN_LC_RUNTIME(LC_COLLATE) | |
84d4ea48 | 1048 | ? ( overloading |
d3fcec1f SP |
1049 | ? (SVCOMPARE_t)S_amagic_cmp_locale |
1050 | : (SVCOMPARE_t)sv_cmp_locale_static) | |
130c5df3 KW |
1051 | : |
1052 | #endif | |
1053 | ( overloading ? (SVCOMPARE_t)S_amagic_cmp : (SVCOMPARE_t)sv_cmp_static)), | |
7b9ef140 | 1054 | sort_flags); |
471178c0 | 1055 | } |
7b9ef140 | 1056 | if ((priv & OPpSORT_REVERSE) != 0) { |
471178c0 NC |
1057 | SV **q = start+max-1; |
1058 | while (start < q) { | |
0bcc34c2 | 1059 | SV * const tmp = *start; |
471178c0 NC |
1060 | *start++ = *q; |
1061 | *q-- = tmp; | |
84d4ea48 JH |
1062 | } |
1063 | } | |
1064 | } | |
84721d61 DM |
1065 | |
1066 | if (av) { | |
1067 | /* copy back result to the array */ | |
1068 | SV** const base = MARK+1; | |
1069 | if (SvMAGICAL(av)) { | |
1070 | for (i = 0; i < max; i++) | |
1071 | base[i] = newSVsv(base[i]); | |
1072 | av_clear(av); | |
1073 | av_extend(av, max); | |
1074 | for (i=0; i < max; i++) { | |
1075 | SV * const sv = base[i]; | |
1076 | SV ** const didstore = av_store(av, i, sv); | |
1077 | if (SvSMAGICAL(sv)) | |
1078 | mg_set(sv); | |
1079 | if (!didstore) | |
1080 | sv_2mortal(sv); | |
1081 | } | |
1082 | } | |
1083 | else { | |
1084 | /* the elements of av are likely to be the same as the | |
1085 | * (non-refcounted) elements on the stack, just in a different | |
1086 | * order. However, its possible that someone's messed with av | |
1087 | * in the meantime. So bump and unbump the relevant refcounts | |
1088 | * first. | |
1089 | */ | |
45c198c1 DM |
1090 | for (i = 0; i < max; i++) { |
1091 | SV *sv = base[i]; | |
1092 | assert(sv); | |
1093 | if (SvREFCNT(sv) > 1) | |
1094 | base[i] = newSVsv(sv); | |
1095 | else | |
1096 | SvREFCNT_inc_simple_void_NN(sv); | |
1097 | } | |
84721d61 DM |
1098 | av_clear(av); |
1099 | if (max > 0) { | |
1100 | av_extend(av, max); | |
1101 | Copy(base, AvARRAY(av), max, SV*); | |
1102 | } | |
1103 | AvFILLp(av) = max - 1; | |
1104 | AvREIFY_off(av); | |
1105 | AvREAL_on(av); | |
1106 | } | |
fe1bc4cf | 1107 | } |
84d4ea48 | 1108 | LEAVE; |
84721d61 | 1109 | PL_stack_sp = ORIGMARK + max; |
84d4ea48 JH |
1110 | return nextop; |
1111 | } | |
1112 | ||
1113 | static I32 | |
31e9e0a3 | 1114 | S_sortcv(pTHX_ SV *const a, SV *const b) |
84d4ea48 | 1115 | { |
901017d6 | 1116 | const I32 oldsaveix = PL_savestack_ix; |
84d4ea48 | 1117 | I32 result; |
ad021bfb | 1118 | PMOP * const pm = PL_curpm; |
a9ea019a | 1119 | COP * const cop = PL_curcop; |
16ada235 | 1120 | SV *olda, *oldb; |
7918f24d NC |
1121 | |
1122 | PERL_ARGS_ASSERT_SORTCV; | |
1123 | ||
16ada235 Z |
1124 | olda = GvSV(PL_firstgv); |
1125 | GvSV(PL_firstgv) = SvREFCNT_inc_simple_NN(a); | |
1126 | SvREFCNT_dec(olda); | |
1127 | oldb = GvSV(PL_secondgv); | |
1128 | GvSV(PL_secondgv) = SvREFCNT_inc_simple_NN(b); | |
1129 | SvREFCNT_dec(oldb); | |
84d4ea48 JH |
1130 | PL_stack_sp = PL_stack_base; |
1131 | PL_op = PL_sortcop; | |
1132 | CALLRUNOPS(aTHX); | |
a9ea019a | 1133 | PL_curcop = cop; |
33411212 DM |
1134 | /* entry zero of a stack is always PL_sv_undef, which |
1135 | * simplifies converting a '()' return into undef in scalar context */ | |
1136 | assert(PL_stack_sp > PL_stack_base || *PL_stack_base == &PL_sv_undef); | |
1137 | result = SvIV(*PL_stack_sp); | |
626ed49c | 1138 | |
53d3542d | 1139 | LEAVE_SCOPE(oldsaveix); |
ad021bfb | 1140 | PL_curpm = pm; |
84d4ea48 JH |
1141 | return result; |
1142 | } | |
1143 | ||
1144 | static I32 | |
31e9e0a3 | 1145 | S_sortcv_stacked(pTHX_ SV *const a, SV *const b) |
84d4ea48 | 1146 | { |
901017d6 | 1147 | const I32 oldsaveix = PL_savestack_ix; |
84d4ea48 | 1148 | I32 result; |
901017d6 | 1149 | AV * const av = GvAV(PL_defgv); |
ad021bfb | 1150 | PMOP * const pm = PL_curpm; |
a9ea019a | 1151 | COP * const cop = PL_curcop; |
84d4ea48 | 1152 | |
7918f24d NC |
1153 | PERL_ARGS_ASSERT_SORTCV_STACKED; |
1154 | ||
8f443ca6 GG |
1155 | if (AvREAL(av)) { |
1156 | av_clear(av); | |
1157 | AvREAL_off(av); | |
1158 | AvREIFY_on(av); | |
1159 | } | |
84d4ea48 | 1160 | if (AvMAX(av) < 1) { |
8f443ca6 | 1161 | SV **ary = AvALLOC(av); |
84d4ea48 JH |
1162 | if (AvARRAY(av) != ary) { |
1163 | AvMAX(av) += AvARRAY(av) - AvALLOC(av); | |
9c6bc640 | 1164 | AvARRAY(av) = ary; |
84d4ea48 JH |
1165 | } |
1166 | if (AvMAX(av) < 1) { | |
84d4ea48 | 1167 | Renew(ary,2,SV*); |
00195859 | 1168 | AvMAX(av) = 1; |
9c6bc640 | 1169 | AvARRAY(av) = ary; |
8f443ca6 | 1170 | AvALLOC(av) = ary; |
84d4ea48 JH |
1171 | } |
1172 | } | |
1173 | AvFILLp(av) = 1; | |
1174 | ||
1175 | AvARRAY(av)[0] = a; | |
1176 | AvARRAY(av)[1] = b; | |
1177 | PL_stack_sp = PL_stack_base; | |
1178 | PL_op = PL_sortcop; | |
1179 | CALLRUNOPS(aTHX); | |
a9ea019a | 1180 | PL_curcop = cop; |
33411212 DM |
1181 | /* entry zero of a stack is always PL_sv_undef, which |
1182 | * simplifies converting a '()' return into undef in scalar context */ | |
1183 | assert(PL_stack_sp > PL_stack_base || *PL_stack_base == &PL_sv_undef); | |
1184 | result = SvIV(*PL_stack_sp); | |
626ed49c | 1185 | |
53d3542d | 1186 | LEAVE_SCOPE(oldsaveix); |
ad021bfb | 1187 | PL_curpm = pm; |
84d4ea48 JH |
1188 | return result; |
1189 | } | |
1190 | ||
1191 | static I32 | |
31e9e0a3 | 1192 | S_sortcv_xsub(pTHX_ SV *const a, SV *const b) |
84d4ea48 | 1193 | { |
20b7effb | 1194 | dSP; |
901017d6 | 1195 | const I32 oldsaveix = PL_savestack_ix; |
ea726b52 | 1196 | CV * const cv=MUTABLE_CV(PL_sortcop); |
84d4ea48 | 1197 | I32 result; |
ad021bfb | 1198 | PMOP * const pm = PL_curpm; |
84d4ea48 | 1199 | |
7918f24d NC |
1200 | PERL_ARGS_ASSERT_SORTCV_XSUB; |
1201 | ||
84d4ea48 JH |
1202 | SP = PL_stack_base; |
1203 | PUSHMARK(SP); | |
1204 | EXTEND(SP, 2); | |
1205 | *++SP = a; | |
1206 | *++SP = b; | |
1207 | PUTBACK; | |
1208 | (void)(*CvXSUB(cv))(aTHX_ cv); | |
33411212 DM |
1209 | /* entry zero of a stack is always PL_sv_undef, which |
1210 | * simplifies converting a '()' return into undef in scalar context */ | |
1211 | assert(PL_stack_sp > PL_stack_base || *PL_stack_base == &PL_sv_undef); | |
84d4ea48 | 1212 | result = SvIV(*PL_stack_sp); |
33411212 | 1213 | |
53d3542d | 1214 | LEAVE_SCOPE(oldsaveix); |
ad021bfb | 1215 | PL_curpm = pm; |
84d4ea48 JH |
1216 | return result; |
1217 | } | |
1218 | ||
1219 | ||
1220 | static I32 | |
31e9e0a3 | 1221 | S_sv_ncmp(pTHX_ SV *const a, SV *const b) |
84d4ea48 | 1222 | { |
427fbfe8 | 1223 | I32 cmp = do_ncmp(a, b); |
7918f24d NC |
1224 | |
1225 | PERL_ARGS_ASSERT_SV_NCMP; | |
1226 | ||
427fbfe8 | 1227 | if (cmp == 2) { |
f3dab52a FC |
1228 | if (ckWARN(WARN_UNINITIALIZED)) report_uninit(NULL); |
1229 | return 0; | |
1230 | } | |
427fbfe8 TC |
1231 | |
1232 | return cmp; | |
84d4ea48 JH |
1233 | } |
1234 | ||
1235 | static I32 | |
31e9e0a3 | 1236 | S_sv_i_ncmp(pTHX_ SV *const a, SV *const b) |
84d4ea48 | 1237 | { |
901017d6 AL |
1238 | const IV iv1 = SvIV(a); |
1239 | const IV iv2 = SvIV(b); | |
7918f24d NC |
1240 | |
1241 | PERL_ARGS_ASSERT_SV_I_NCMP; | |
1242 | ||
84d4ea48 JH |
1243 | return iv1 < iv2 ? -1 : iv1 > iv2 ? 1 : 0; |
1244 | } | |
901017d6 AL |
1245 | |
1246 | #define tryCALL_AMAGICbin(left,right,meth) \ | |
79a8d529 | 1247 | (SvAMAGIC(left)||SvAMAGIC(right)) \ |
31d632c3 | 1248 | ? amagic_call(left, right, meth, 0) \ |
a0714e2c | 1249 | : NULL; |
84d4ea48 | 1250 | |
659c4b96 | 1251 | #define SORT_NORMAL_RETURN_VALUE(val) (((val) > 0) ? 1 : ((val) ? -1 : 0)) |
eeb9de02 | 1252 | |
84d4ea48 | 1253 | static I32 |
5aaab254 | 1254 | S_amagic_ncmp(pTHX_ SV *const a, SV *const b) |
84d4ea48 | 1255 | { |
31d632c3 | 1256 | SV * const tmpsv = tryCALL_AMAGICbin(a,b,ncmp_amg); |
7918f24d NC |
1257 | |
1258 | PERL_ARGS_ASSERT_AMAGIC_NCMP; | |
1259 | ||
84d4ea48 | 1260 | if (tmpsv) { |
84d4ea48 | 1261 | if (SvIOK(tmpsv)) { |
901017d6 | 1262 | const I32 i = SvIVX(tmpsv); |
eeb9de02 | 1263 | return SORT_NORMAL_RETURN_VALUE(i); |
84d4ea48 | 1264 | } |
901017d6 AL |
1265 | else { |
1266 | const NV d = SvNV(tmpsv); | |
eeb9de02 | 1267 | return SORT_NORMAL_RETURN_VALUE(d); |
901017d6 | 1268 | } |
84d4ea48 | 1269 | } |
f0f5dc9d | 1270 | return S_sv_ncmp(aTHX_ a, b); |
84d4ea48 JH |
1271 | } |
1272 | ||
1273 | static I32 | |
5aaab254 | 1274 | S_amagic_i_ncmp(pTHX_ SV *const a, SV *const b) |
84d4ea48 | 1275 | { |
31d632c3 | 1276 | SV * const tmpsv = tryCALL_AMAGICbin(a,b,ncmp_amg); |
7918f24d NC |
1277 | |
1278 | PERL_ARGS_ASSERT_AMAGIC_I_NCMP; | |
1279 | ||
84d4ea48 | 1280 | if (tmpsv) { |
84d4ea48 | 1281 | if (SvIOK(tmpsv)) { |
901017d6 | 1282 | const I32 i = SvIVX(tmpsv); |
eeb9de02 | 1283 | return SORT_NORMAL_RETURN_VALUE(i); |
84d4ea48 | 1284 | } |
901017d6 AL |
1285 | else { |
1286 | const NV d = SvNV(tmpsv); | |
eeb9de02 | 1287 | return SORT_NORMAL_RETURN_VALUE(d); |
901017d6 | 1288 | } |
84d4ea48 | 1289 | } |
f0f5dc9d | 1290 | return S_sv_i_ncmp(aTHX_ a, b); |
84d4ea48 JH |
1291 | } |
1292 | ||
1293 | static I32 | |
5aaab254 | 1294 | S_amagic_cmp(pTHX_ SV *const str1, SV *const str2) |
84d4ea48 | 1295 | { |
31d632c3 | 1296 | SV * const tmpsv = tryCALL_AMAGICbin(str1,str2,scmp_amg); |
7918f24d NC |
1297 | |
1298 | PERL_ARGS_ASSERT_AMAGIC_CMP; | |
1299 | ||
84d4ea48 | 1300 | if (tmpsv) { |
84d4ea48 | 1301 | if (SvIOK(tmpsv)) { |
901017d6 | 1302 | const I32 i = SvIVX(tmpsv); |
eeb9de02 | 1303 | return SORT_NORMAL_RETURN_VALUE(i); |
84d4ea48 | 1304 | } |
901017d6 AL |
1305 | else { |
1306 | const NV d = SvNV(tmpsv); | |
eeb9de02 | 1307 | return SORT_NORMAL_RETURN_VALUE(d); |
901017d6 | 1308 | } |
84d4ea48 JH |
1309 | } |
1310 | return sv_cmp(str1, str2); | |
1311 | } | |
1312 | ||
91191cf7 KW |
1313 | #ifdef USE_LOCALE_COLLATE |
1314 | ||
84d4ea48 | 1315 | static I32 |
5aaab254 | 1316 | S_amagic_cmp_locale(pTHX_ SV *const str1, SV *const str2) |
84d4ea48 | 1317 | { |
31d632c3 | 1318 | SV * const tmpsv = tryCALL_AMAGICbin(str1,str2,scmp_amg); |
7918f24d NC |
1319 | |
1320 | PERL_ARGS_ASSERT_AMAGIC_CMP_LOCALE; | |
1321 | ||
84d4ea48 | 1322 | if (tmpsv) { |
84d4ea48 | 1323 | if (SvIOK(tmpsv)) { |
901017d6 | 1324 | const I32 i = SvIVX(tmpsv); |
eeb9de02 | 1325 | return SORT_NORMAL_RETURN_VALUE(i); |
84d4ea48 | 1326 | } |
901017d6 AL |
1327 | else { |
1328 | const NV d = SvNV(tmpsv); | |
eeb9de02 | 1329 | return SORT_NORMAL_RETURN_VALUE(d); |
901017d6 | 1330 | } |
84d4ea48 JH |
1331 | } |
1332 | return sv_cmp_locale(str1, str2); | |
1333 | } | |
241d1a3b | 1334 | |
91191cf7 KW |
1335 | #endif |
1336 | ||
241d1a3b | 1337 | /* |
14d04a33 | 1338 | * ex: set ts=8 sts=4 sw=4 et: |
37442d52 | 1339 | */ |