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 | ||
c53fc8a6 | 33 | #ifndef SMALLSORT |
7ea738a9 | 34 | #define SMALLSORT (200) |
c53fc8a6 JH |
35 | #endif |
36 | ||
23a85c22 | 37 | /* Flags for sortsv_flags */ |
044d25c7 TK |
38 | #define SORTf_STABLE 1 |
39 | #define SORTf_UNSTABLE 2 | |
7b9ef140 | 40 | |
84d4ea48 JH |
41 | /* |
42 | * The mergesort implementation is by Peter M. Mcilroy <pmcilroy@lucent.com>. | |
43 | * | |
44 | * The original code was written in conjunction with BSD Computer Software | |
45 | * Research Group at University of California, Berkeley. | |
46 | * | |
393db44d JL |
47 | * See also: "Optimistic Sorting and Information Theoretic Complexity" |
48 | * Peter McIlroy | |
49 | * SODA (Fourth Annual ACM-SIAM Symposium on Discrete Algorithms), | |
50 | * pp 467-474, Austin, Texas, 25-27 January 1993. | |
84d4ea48 | 51 | * |
393db44d | 52 | * The integration to Perl is by John P. Linderman <jpl.jpl@gmail.com>. |
84d4ea48 JH |
53 | * |
54 | * The code can be distributed under the same terms as Perl itself. | |
55 | * | |
56 | */ | |
57 | ||
84d4ea48 | 58 | |
7ea738a9 TK |
59 | typedef char * aptr; /* pointer for arithmetic on sizes */ |
60 | typedef SV * gptr; /* pointers in our lists */ | |
84d4ea48 JH |
61 | |
62 | /* Binary merge internal sort, with a few special mods | |
63 | ** for the special perl environment it now finds itself in. | |
64 | ** | |
65 | ** Things that were once options have been hotwired | |
66 | ** to values suitable for this use. In particular, we'll always | |
67 | ** initialize looking for natural runs, we'll always produce stable | |
68 | ** output, and we'll always do Peter McIlroy's binary merge. | |
69 | */ | |
70 | ||
71 | /* Pointer types for arithmetic and storage and convenience casts */ | |
72 | ||
7ea738a9 TK |
73 | #define APTR(P) ((aptr)(P)) |
74 | #define GPTP(P) ((gptr *)(P)) | |
84d4ea48 JH |
75 | #define GPPP(P) ((gptr **)(P)) |
76 | ||
77 | ||
78 | /* byte offset from pointer P to (larger) pointer Q */ | |
7ea738a9 | 79 | #define BYTEOFF(P, Q) (APTR(Q) - APTR(P)) |
84d4ea48 JH |
80 | |
81 | #define PSIZE sizeof(gptr) | |
82 | ||
83 | /* If PSIZE is power of 2, make PSHIFT that power, if that helps */ | |
84 | ||
7ea738a9 TK |
85 | #ifdef PSHIFT |
86 | #define PNELEM(P, Q) (BYTEOFF(P,Q) >> (PSHIFT)) | |
87 | #define PNBYTE(N) ((N) << (PSHIFT)) | |
88 | #define PINDEX(P, N) (GPTP(APTR(P) + PNBYTE(N))) | |
84d4ea48 JH |
89 | #else |
90 | /* Leave optimization to compiler */ | |
7ea738a9 TK |
91 | #define PNELEM(P, Q) (GPTP(Q) - GPTP(P)) |
92 | #define PNBYTE(N) ((N) * (PSIZE)) | |
93 | #define PINDEX(P, N) (GPTP(P) + (N)) | |
84d4ea48 JH |
94 | #endif |
95 | ||
96 | /* Pointer into other corresponding to pointer into this */ | |
7ea738a9 | 97 | #define POTHER(P, THIS, OTHER) GPTP(APTR(OTHER) + BYTEOFF(THIS,P)) |
84d4ea48 JH |
98 | |
99 | #define FROMTOUPTO(src, dst, lim) do *dst++ = *src++; while(src<lim) | |
100 | ||
101 | ||
486ec47a | 102 | /* Runs are identified by a pointer in the auxiliary list. |
84d4ea48 JH |
103 | ** The pointer is at the start of the list, |
104 | ** and it points to the start of the next list. | |
105 | ** NEXT is used as an lvalue, too. | |
106 | */ | |
107 | ||
7ea738a9 | 108 | #define NEXT(P) (*GPPP(P)) |
84d4ea48 JH |
109 | |
110 | ||
111 | /* PTHRESH is the minimum number of pairs with the same sense to justify | |
112 | ** checking for a run and extending it. Note that PTHRESH counts PAIRS, | |
113 | ** not just elements, so PTHRESH == 8 means a run of 16. | |
114 | */ | |
115 | ||
7ea738a9 | 116 | #define PTHRESH (8) |
84d4ea48 JH |
117 | |
118 | /* RTHRESH is the number of elements in a run that must compare low | |
119 | ** to the low element from the opposing run before we justify | |
120 | ** doing a binary rampup instead of single stepping. | |
121 | ** In random input, N in a row low should only happen with | |
122 | ** probability 2^(1-N), so we can risk that we are dealing | |
123 | ** with orderly input without paying much when we aren't. | |
124 | */ | |
125 | ||
126 | #define RTHRESH (6) | |
127 | ||
128 | ||
129 | /* | |
130 | ** Overview of algorithm and variables. | |
131 | ** The array of elements at list1 will be organized into runs of length 2, | |
132 | ** or runs of length >= 2 * PTHRESH. We only try to form long runs when | |
133 | ** PTHRESH adjacent pairs compare in the same way, suggesting overall order. | |
134 | ** | |
135 | ** Unless otherwise specified, pair pointers address the first of two elements. | |
136 | ** | |
a0288114 AL |
137 | ** b and b+1 are a pair that compare with sense "sense". |
138 | ** b is the "bottom" of adjacent pairs that might form a longer run. | |
84d4ea48 JH |
139 | ** |
140 | ** p2 parallels b in the list2 array, where runs are defined by | |
141 | ** a pointer chain. | |
142 | ** | |
a0288114 | 143 | ** t represents the "top" of the adjacent pairs that might extend |
84d4ea48 JH |
144 | ** the run beginning at b. Usually, t addresses a pair |
145 | ** that compares with opposite sense from (b,b+1). | |
146 | ** However, it may also address a singleton element at the end of list1, | |
a0288114 | 147 | ** or it may be equal to "last", the first element beyond list1. |
84d4ea48 JH |
148 | ** |
149 | ** r addresses the Nth pair following b. If this would be beyond t, | |
150 | ** we back it off to t. Only when r is less than t do we consider the | |
151 | ** run long enough to consider checking. | |
152 | ** | |
153 | ** q addresses a pair such that the pairs at b through q already form a run. | |
154 | ** Often, q will equal b, indicating we only are sure of the pair itself. | |
155 | ** However, a search on the previous cycle may have revealed a longer run, | |
156 | ** so q may be greater than b. | |
157 | ** | |
158 | ** p is used to work back from a candidate r, trying to reach q, | |
159 | ** which would mean b through r would be a run. If we discover such a run, | |
160 | ** we start q at r and try to push it further towards t. | |
161 | ** If b through r is NOT a run, we detect the wrong order at (p-1,p). | |
162 | ** In any event, after the check (if any), we have two main cases. | |
163 | ** | |
164 | ** 1) Short run. b <= q < p <= r <= t. | |
7ea738a9 TK |
165 | ** b through q is a run (perhaps trivial) |
166 | ** q through p are uninteresting pairs | |
167 | ** p through r is a run | |
84d4ea48 JH |
168 | ** |
169 | ** 2) Long run. b < r <= q < t. | |
7ea738a9 | 170 | ** b through q is a run (of length >= 2 * PTHRESH) |
84d4ea48 JH |
171 | ** |
172 | ** Note that degenerate cases are not only possible, but likely. | |
173 | ** For example, if the pair following b compares with opposite sense, | |
174 | ** then b == q < p == r == t. | |
175 | */ | |
176 | ||
177 | ||
044d25c7 | 178 | PERL_STATIC_FORCE_INLINE IV __attribute__always_inline__ |
d4c19fe8 | 179 | dynprep(pTHX_ gptr *list1, gptr *list2, size_t nmemb, const SVCOMPARE_t cmp) |
84d4ea48 | 180 | { |
957d8989 | 181 | I32 sense; |
eb578fdb KW |
182 | gptr *b, *p, *q, *t, *p2; |
183 | gptr *last, *r; | |
957d8989 | 184 | IV runs = 0; |
84d4ea48 JH |
185 | |
186 | b = list1; | |
187 | last = PINDEX(b, nmemb); | |
188 | sense = (cmp(aTHX_ *b, *(b+1)) > 0); | |
189 | for (p2 = list2; b < last; ) { | |
7ea738a9 TK |
190 | /* We just started, or just reversed sense. |
191 | ** Set t at end of pairs with the prevailing sense. | |
192 | */ | |
193 | for (p = b+2, t = p; ++p < last; t = ++p) { | |
194 | if ((cmp(aTHX_ *t, *p) > 0) != sense) break; | |
195 | } | |
196 | q = b; | |
197 | /* Having laid out the playing field, look for long runs */ | |
198 | do { | |
199 | p = r = b + (2 * PTHRESH); | |
200 | if (r >= t) p = r = t; /* too short to care about */ | |
201 | else { | |
202 | while (((cmp(aTHX_ *(p-1), *p) > 0) == sense) && | |
203 | ((p -= 2) > q)) {} | |
204 | if (p <= q) { | |
205 | /* b through r is a (long) run. | |
206 | ** Extend it as far as possible. | |
207 | */ | |
208 | p = q = r; | |
209 | while (((p += 2) < t) && | |
210 | ((cmp(aTHX_ *(p-1), *p) > 0) == sense)) q = p; | |
211 | r = p = q + 2; /* no simple pairs, no after-run */ | |
212 | } | |
213 | } | |
214 | if (q > b) { /* run of greater than 2 at b */ | |
215 | gptr *savep = p; | |
216 | ||
217 | p = q += 2; | |
218 | /* pick up singleton, if possible */ | |
219 | if ((p == t) && | |
220 | ((t + 1) == last) && | |
221 | ((cmp(aTHX_ *(p-1), *p) > 0) == sense)) | |
222 | savep = r = p = q = last; | |
223 | p2 = NEXT(p2) = p2 + (p - b); ++runs; | |
224 | if (sense) | |
225 | while (b < --p) { | |
226 | const gptr c = *b; | |
227 | *b++ = *p; | |
228 | *p = c; | |
229 | } | |
230 | p = savep; | |
231 | } | |
232 | while (q < p) { /* simple pairs */ | |
233 | p2 = NEXT(p2) = p2 + 2; ++runs; | |
234 | if (sense) { | |
235 | const gptr c = *q++; | |
236 | *(q-1) = *q; | |
237 | *q++ = c; | |
238 | } else q += 2; | |
239 | } | |
240 | if (((b = p) == t) && ((t+1) == last)) { | |
241 | NEXT(p2) = p2 + 1; ++runs; | |
242 | b++; | |
243 | } | |
244 | q = r; | |
245 | } while (b < t); | |
246 | sense = !sense; | |
84d4ea48 | 247 | } |
957d8989 | 248 | return runs; |
84d4ea48 JH |
249 | } |
250 | ||
251 | ||
3fe0b9a9 | 252 | /* The original merge sort, in use since 5.7, was as fast as, or faster than, |
957d8989 | 253 | * qsort on many platforms, but slower than qsort, conspicuously so, |
3fe0b9a9 | 254 | * on others. The most likely explanation was platform-specific |
957d8989 JL |
255 | * differences in cache sizes and relative speeds. |
256 | * | |
257 | * The quicksort divide-and-conquer algorithm guarantees that, as the | |
258 | * problem is subdivided into smaller and smaller parts, the parts | |
259 | * fit into smaller (and faster) caches. So it doesn't matter how | |
260 | * many levels of cache exist, quicksort will "find" them, and, | |
e62b3022 | 261 | * as long as smaller is faster, take advantage of them. |
957d8989 | 262 | * |
3fe0b9a9 | 263 | * By contrast, consider how the original mergesort algorithm worked. |
957d8989 JL |
264 | * Suppose we have five runs (each typically of length 2 after dynprep). |
265 | * | |
266 | * pass base aux | |
267 | * 0 1 2 3 4 5 | |
268 | * 1 12 34 5 | |
269 | * 2 1234 5 | |
270 | * 3 12345 | |
271 | * 4 12345 | |
272 | * | |
273 | * Adjacent pairs are merged in "grand sweeps" through the input. | |
274 | * This means, on pass 1, the records in runs 1 and 2 aren't revisited until | |
275 | * runs 3 and 4 are merged and the runs from run 5 have been copied. | |
276 | * The only cache that matters is one large enough to hold *all* the input. | |
277 | * On some platforms, this may be many times slower than smaller caches. | |
278 | * | |
279 | * The following pseudo-code uses the same basic merge algorithm, | |
280 | * but in a divide-and-conquer way. | |
281 | * | |
282 | * # merge $runs runs at offset $offset of list $list1 into $list2. | |
283 | * # all unmerged runs ($runs == 1) originate in list $base. | |
284 | * sub mgsort2 { | |
285 | * my ($offset, $runs, $base, $list1, $list2) = @_; | |
286 | * | |
287 | * if ($runs == 1) { | |
288 | * if ($list1 is $base) copy run to $list2 | |
289 | * return offset of end of list (or copy) | |
290 | * } else { | |
291 | * $off2 = mgsort2($offset, $runs-($runs/2), $base, $list2, $list1) | |
292 | * mgsort2($off2, $runs/2, $base, $list2, $list1) | |
293 | * merge the adjacent runs at $offset of $list1 into $list2 | |
294 | * return the offset of the end of the merged runs | |
295 | * } | |
296 | * } | |
297 | * mgsort2(0, $runs, $base, $aux, $base); | |
298 | * | |
299 | * For our 5 runs, the tree of calls looks like | |
300 | * | |
301 | * 5 | |
302 | * 3 2 | |
303 | * 2 1 1 1 | |
304 | * 1 1 | |
305 | * | |
306 | * 1 2 3 4 5 | |
307 | * | |
308 | * and the corresponding activity looks like | |
309 | * | |
310 | * copy runs 1 and 2 from base to aux | |
311 | * merge runs 1 and 2 from aux to base | |
312 | * (run 3 is where it belongs, no copy needed) | |
313 | * merge runs 12 and 3 from base to aux | |
314 | * (runs 4 and 5 are where they belong, no copy needed) | |
315 | * merge runs 4 and 5 from base to aux | |
316 | * merge runs 123 and 45 from aux to base | |
317 | * | |
318 | * Note that we merge runs 1 and 2 immediately after copying them, | |
319 | * while they are still likely to be in fast cache. Similarly, | |
320 | * run 3 is merged with run 12 while it still may be lingering in cache. | |
321 | * This implementation should therefore enjoy much of the cache-friendly | |
322 | * behavior that quicksort does. In addition, it does less copying | |
323 | * than the original mergesort implementation (only runs 1 and 2 are copied) | |
324 | * and the "balancing" of merges is better (merged runs comprise more nearly | |
325 | * equal numbers of original runs). | |
326 | * | |
327 | * The actual cache-friendly implementation will use a pseudo-stack | |
328 | * to avoid recursion, and will unroll processing of runs of length 2, | |
329 | * but it is otherwise similar to the recursive implementation. | |
957d8989 JL |
330 | */ |
331 | ||
332 | typedef struct { | |
7ea738a9 TK |
333 | IV offset; /* offset of 1st of 2 runs at this level */ |
334 | IV runs; /* how many runs must be combined into 1 */ | |
335 | } off_runs; /* pseudo-stack element */ | |
957d8989 | 336 | |
044d25c7 TK |
337 | PERL_STATIC_FORCE_INLINE void |
338 | S_sortsv_flags_impl(pTHX_ gptr *base, size_t nmemb, SVCOMPARE_t cmp, U32 flags) | |
957d8989 | 339 | { |
551405c4 | 340 | IV i, run, offset; |
957d8989 | 341 | I32 sense, level; |
eb578fdb | 342 | gptr *f1, *f2, *t, *b, *p; |
957d8989 | 343 | int iwhich; |
551405c4 | 344 | gptr *aux; |
957d8989 JL |
345 | gptr *p1; |
346 | gptr small[SMALLSORT]; | |
347 | gptr *which[3]; | |
348 | off_runs stack[60], *stackp; | |
349 | ||
aa4119bb | 350 | PERL_UNUSED_ARG(flags); |
044d25c7 | 351 | PERL_ARGS_ASSERT_SORTSV_FLAGS_IMPL; |
7ea738a9 | 352 | if (nmemb <= 1) return; /* sorted trivially */ |
6c3fb703 | 353 | |
7ea738a9 TK |
354 | if (nmemb <= SMALLSORT) aux = small; /* use stack for aux array */ |
355 | else { Newx(aux,nmemb,gptr); } /* allocate auxiliary array */ | |
957d8989 JL |
356 | level = 0; |
357 | stackp = stack; | |
358 | stackp->runs = dynprep(aTHX_ base, aux, nmemb, cmp); | |
359 | stackp->offset = offset = 0; | |
360 | which[0] = which[2] = base; | |
361 | which[1] = aux; | |
362 | for (;;) { | |
7ea738a9 TK |
363 | /* On levels where both runs have be constructed (stackp->runs == 0), |
364 | * merge them, and note the offset of their end, in case the offset | |
365 | * is needed at the next level up. Hop up a level, and, | |
366 | * as long as stackp->runs is 0, keep merging. | |
367 | */ | |
368 | IV runs = stackp->runs; | |
369 | if (runs == 0) { | |
370 | gptr *list1, *list2; | |
371 | iwhich = level & 1; | |
372 | list1 = which[iwhich]; /* area where runs are now */ | |
373 | list2 = which[++iwhich]; /* area for merged runs */ | |
374 | do { | |
375 | gptr *l1, *l2, *tp2; | |
376 | offset = stackp->offset; | |
377 | f1 = p1 = list1 + offset; /* start of first run */ | |
378 | p = tp2 = list2 + offset; /* where merged run will go */ | |
379 | t = NEXT(p); /* where first run ends */ | |
380 | f2 = l1 = POTHER(t, list2, list1); /* ... on the other side */ | |
381 | t = NEXT(t); /* where second runs ends */ | |
382 | l2 = POTHER(t, list2, list1); /* ... on the other side */ | |
383 | offset = PNELEM(list2, t); | |
384 | while (f1 < l1 && f2 < l2) { | |
385 | /* If head 1 is larger than head 2, find ALL the elements | |
386 | ** in list 2 strictly less than head1, write them all, | |
387 | ** then head 1. Then compare the new heads, and repeat, | |
388 | ** until one or both lists are exhausted. | |
389 | ** | |
390 | ** In all comparisons (after establishing | |
391 | ** which head to merge) the item to merge | |
392 | ** (at pointer q) is the first operand of | |
393 | ** the comparison. When we want to know | |
394 | ** if "q is strictly less than the other", | |
395 | ** we can't just do | |
396 | ** cmp(q, other) < 0 | |
397 | ** because stability demands that we treat equality | |
398 | ** as high when q comes from l2, and as low when | |
399 | ** q was from l1. So we ask the question by doing | |
400 | ** cmp(q, other) <= sense | |
401 | ** and make sense == 0 when equality should look low, | |
402 | ** and -1 when equality should look high. | |
403 | */ | |
404 | ||
405 | gptr *q; | |
406 | if (cmp(aTHX_ *f1, *f2) <= 0) { | |
407 | q = f2; b = f1; t = l1; | |
408 | sense = -1; | |
409 | } else { | |
410 | q = f1; b = f2; t = l2; | |
411 | sense = 0; | |
412 | } | |
413 | ||
414 | ||
415 | /* ramp up | |
416 | ** | |
417 | ** Leave t at something strictly | |
418 | ** greater than q (or at the end of the list), | |
419 | ** and b at something strictly less than q. | |
420 | */ | |
421 | for (i = 1, run = 0 ;;) { | |
422 | if ((p = PINDEX(b, i)) >= t) { | |
423 | /* off the end */ | |
424 | if (((p = PINDEX(t, -1)) > b) && | |
425 | (cmp(aTHX_ *q, *p) <= sense)) | |
426 | t = p; | |
427 | else b = p; | |
428 | break; | |
429 | } else if (cmp(aTHX_ *q, *p) <= sense) { | |
430 | t = p; | |
431 | break; | |
432 | } else b = p; | |
433 | if (++run >= RTHRESH) i += i; | |
434 | } | |
435 | ||
436 | ||
437 | /* q is known to follow b and must be inserted before t. | |
438 | ** Increment b, so the range of possibilities is [b,t). | |
439 | ** Round binary split down, to favor early appearance. | |
440 | ** Adjust b and t until q belongs just before t. | |
441 | */ | |
442 | ||
443 | b++; | |
444 | while (b < t) { | |
445 | p = PINDEX(b, (PNELEM(b, t) - 1) / 2); | |
446 | if (cmp(aTHX_ *q, *p) <= sense) { | |
447 | t = p; | |
448 | } else b = p + 1; | |
449 | } | |
450 | ||
451 | ||
452 | /* Copy all the strictly low elements */ | |
453 | ||
454 | if (q == f1) { | |
455 | FROMTOUPTO(f2, tp2, t); | |
456 | *tp2++ = *f1++; | |
457 | } else { | |
458 | FROMTOUPTO(f1, tp2, t); | |
459 | *tp2++ = *f2++; | |
460 | } | |
461 | } | |
462 | ||
463 | ||
464 | /* Run out remaining list */ | |
465 | if (f1 == l1) { | |
466 | if (f2 < l2) FROMTOUPTO(f2, tp2, l2); | |
467 | } else FROMTOUPTO(f1, tp2, l1); | |
468 | p1 = NEXT(p1) = POTHER(tp2, list2, list1); | |
469 | ||
470 | if (--level == 0) goto done; | |
471 | --stackp; | |
472 | t = list1; list1 = list2; list2 = t; /* swap lists */ | |
473 | } while ((runs = stackp->runs) == 0); | |
474 | } | |
475 | ||
476 | ||
477 | stackp->runs = 0; /* current run will finish level */ | |
478 | /* While there are more than 2 runs remaining, | |
479 | * turn them into exactly 2 runs (at the "other" level), | |
480 | * each made up of approximately half the runs. | |
481 | * Stack the second half for later processing, | |
482 | * and set about producing the first half now. | |
483 | */ | |
484 | while (runs > 2) { | |
485 | ++level; | |
486 | ++stackp; | |
487 | stackp->offset = offset; | |
488 | runs -= stackp->runs = runs / 2; | |
489 | } | |
490 | /* We must construct a single run from 1 or 2 runs. | |
491 | * All the original runs are in which[0] == base. | |
492 | * The run we construct must end up in which[level&1]. | |
493 | */ | |
494 | iwhich = level & 1; | |
495 | if (runs == 1) { | |
496 | /* Constructing a single run from a single run. | |
497 | * If it's where it belongs already, there's nothing to do. | |
498 | * Otherwise, copy it to where it belongs. | |
499 | * A run of 1 is either a singleton at level 0, | |
500 | * or the second half of a split 3. In neither event | |
501 | * is it necessary to set offset. It will be set by the merge | |
502 | * that immediately follows. | |
503 | */ | |
504 | if (iwhich) { /* Belongs in aux, currently in base */ | |
505 | f1 = b = PINDEX(base, offset); /* where list starts */ | |
506 | f2 = PINDEX(aux, offset); /* where list goes */ | |
507 | t = NEXT(f2); /* where list will end */ | |
508 | offset = PNELEM(aux, t); /* offset thereof */ | |
509 | t = PINDEX(base, offset); /* where it currently ends */ | |
510 | FROMTOUPTO(f1, f2, t); /* copy */ | |
511 | NEXT(b) = t; /* set up parallel pointer */ | |
512 | } else if (level == 0) goto done; /* single run at level 0 */ | |
513 | } else { | |
514 | /* Constructing a single run from two runs. | |
515 | * The merge code at the top will do that. | |
516 | * We need only make sure the two runs are in the "other" array, | |
517 | * so they'll end up in the correct array after the merge. | |
518 | */ | |
519 | ++level; | |
520 | ++stackp; | |
521 | stackp->offset = offset; | |
522 | stackp->runs = 0; /* take care of both runs, trigger merge */ | |
523 | if (!iwhich) { /* Merged runs belong in aux, copy 1st */ | |
524 | f1 = b = PINDEX(base, offset); /* where first run starts */ | |
525 | f2 = PINDEX(aux, offset); /* where it will be copied */ | |
526 | t = NEXT(f2); /* where first run will end */ | |
527 | offset = PNELEM(aux, t); /* offset thereof */ | |
528 | p = PINDEX(base, offset); /* end of first run */ | |
529 | t = NEXT(t); /* where second run will end */ | |
530 | t = PINDEX(base, PNELEM(aux, t)); /* where it now ends */ | |
531 | FROMTOUPTO(f1, f2, t); /* copy both runs */ | |
532 | NEXT(b) = p; /* paralleled pointer for 1st */ | |
533 | NEXT(p) = t; /* ... and for second */ | |
534 | } | |
535 | } | |
957d8989 | 536 | } |
7b52d656 | 537 | done: |
7ea738a9 | 538 | if (aux != small) Safefree(aux); /* free iff allocated */ |
044d25c7 | 539 | |
957d8989 JL |
540 | return; |
541 | } | |
542 | ||
84d4ea48 | 543 | /* |
51b56f5c | 544 | =for apidoc_section SV Handling |
044d25c7 TK |
545 | |
546 | =for apidoc sortsv_flags | |
547 | ||
548 | In-place sort an array of SV pointers with the given comparison routine, | |
549 | with various SORTf_* flag options. | |
550 | ||
551 | =cut | |
552 | */ | |
553 | void | |
554 | Perl_sortsv_flags(pTHX_ gptr *base, size_t nmemb, SVCOMPARE_t cmp, U32 flags) | |
555 | { | |
556 | PERL_ARGS_ASSERT_SORTSV_FLAGS; | |
557 | ||
558 | sortsv_flags_impl(base, nmemb, cmp, flags); | |
559 | } | |
560 | ||
561 | /* | |
562 | * Each of sortsv_* functions contains an inlined copy of | |
563 | * sortsv_flags_impl() with an inlined comparator. Basically, we are | |
564 | * emulating C++ templates by using __attribute__((always_inline)). | |
565 | * | |
566 | * The purpose of that is to avoid the function call overhead inside | |
567 | * the sorting routine, which calls the comparison function multiple | |
568 | * times per sorted item. | |
569 | */ | |
570 | ||
571 | static void | |
572 | sortsv_amagic_i_ncmp(pTHX_ gptr *base, size_t nmemb, U32 flags) | |
573 | { | |
574 | sortsv_flags_impl(base, nmemb, S_amagic_i_ncmp, flags); | |
575 | } | |
576 | ||
577 | static void | |
578 | sortsv_amagic_i_ncmp_desc(pTHX_ gptr *base, size_t nmemb, U32 flags) | |
579 | { | |
580 | sortsv_flags_impl(base, nmemb, S_amagic_i_ncmp_desc, flags); | |
581 | } | |
582 | ||
583 | static void | |
584 | sortsv_i_ncmp(pTHX_ gptr *base, size_t nmemb, U32 flags) | |
585 | { | |
586 | sortsv_flags_impl(base, nmemb, S_sv_i_ncmp, flags); | |
587 | } | |
588 | ||
589 | static void | |
590 | sortsv_i_ncmp_desc(pTHX_ gptr *base, size_t nmemb, U32 flags) | |
591 | { | |
592 | sortsv_flags_impl(base, nmemb, S_sv_i_ncmp_desc, flags); | |
593 | } | |
594 | ||
595 | static void | |
596 | sortsv_amagic_ncmp(pTHX_ gptr *base, size_t nmemb, U32 flags) | |
597 | { | |
598 | sortsv_flags_impl(base, nmemb, S_amagic_ncmp, flags); | |
599 | } | |
600 | ||
601 | static void | |
602 | sortsv_amagic_ncmp_desc(pTHX_ gptr *base, size_t nmemb, U32 flags) | |
603 | { | |
604 | sortsv_flags_impl(base, nmemb, S_amagic_ncmp_desc, flags); | |
605 | } | |
606 | ||
607 | static void | |
608 | sortsv_ncmp(pTHX_ gptr *base, size_t nmemb, U32 flags) | |
609 | { | |
610 | sortsv_flags_impl(base, nmemb, S_sv_ncmp, flags); | |
611 | } | |
612 | ||
613 | static void | |
614 | sortsv_ncmp_desc(pTHX_ gptr *base, size_t nmemb, U32 flags) | |
615 | { | |
616 | sortsv_flags_impl(base, nmemb, S_sv_ncmp_desc, flags); | |
617 | } | |
618 | ||
619 | static void | |
620 | sortsv_amagic_cmp(pTHX_ gptr *base, size_t nmemb, U32 flags) | |
621 | { | |
622 | sortsv_flags_impl(base, nmemb, S_amagic_cmp, flags); | |
623 | } | |
624 | ||
625 | static void | |
626 | sortsv_amagic_cmp_desc(pTHX_ gptr *base, size_t nmemb, U32 flags) | |
627 | { | |
628 | sortsv_flags_impl(base, nmemb, S_amagic_cmp_desc, flags); | |
629 | } | |
630 | ||
631 | static void | |
632 | sortsv_cmp(pTHX_ gptr *base, size_t nmemb, U32 flags) | |
633 | { | |
634 | sortsv_flags_impl(base, nmemb, Perl_sv_cmp, flags); | |
635 | } | |
636 | ||
637 | static void | |
638 | sortsv_cmp_desc(pTHX_ gptr *base, size_t nmemb, U32 flags) | |
639 | { | |
640 | sortsv_flags_impl(base, nmemb, S_cmp_desc, flags); | |
641 | } | |
642 | ||
643 | #ifdef USE_LOCALE_COLLATE | |
644 | ||
645 | static void | |
646 | sortsv_amagic_cmp_locale(pTHX_ gptr *base, size_t nmemb, U32 flags) | |
647 | { | |
648 | sortsv_flags_impl(base, nmemb, S_amagic_cmp_locale, flags); | |
649 | } | |
650 | ||
651 | static void | |
652 | sortsv_amagic_cmp_locale_desc(pTHX_ gptr *base, size_t nmemb, U32 flags) | |
653 | { | |
654 | sortsv_flags_impl(base, nmemb, S_amagic_cmp_locale_desc, flags); | |
655 | } | |
656 | ||
657 | static void | |
658 | sortsv_cmp_locale(pTHX_ gptr *base, size_t nmemb, U32 flags) | |
659 | { | |
660 | sortsv_flags_impl(base, nmemb, Perl_sv_cmp_locale, flags); | |
661 | } | |
662 | ||
663 | static void | |
664 | sortsv_cmp_locale_desc(pTHX_ gptr *base, size_t nmemb, U32 flags) | |
665 | { | |
666 | sortsv_flags_impl(base, nmemb, S_cmp_locale_desc, flags); | |
667 | } | |
668 | ||
669 | #endif | |
670 | ||
671 | /* | |
ccfc67b7 | 672 | |
84d4ea48 JH |
673 | =for apidoc sortsv |
674 | ||
8f5d5a51 | 675 | In-place sort an array of SV pointers with the given comparison routine. |
84d4ea48 | 676 | |
796b6530 | 677 | Currently this always uses mergesort. See C<L</sortsv_flags>> for a more |
7b9ef140 | 678 | flexible routine. |
78210658 | 679 | |
84d4ea48 JH |
680 | =cut |
681 | */ | |
4eb872f6 | 682 | |
84d4ea48 JH |
683 | void |
684 | Perl_sortsv(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp) | |
685 | { | |
7918f24d NC |
686 | PERL_ARGS_ASSERT_SORTSV; |
687 | ||
7b9ef140 | 688 | sortsv_flags(array, nmemb, cmp, 0); |
6c3fb703 NC |
689 | } |
690 | ||
4d562308 SF |
691 | #define SvNSIOK(sv) ((SvFLAGS(sv) & SVf_NOK) || ((SvFLAGS(sv) & (SVf_IOK|SVf_IVisUV)) == SVf_IOK)) |
692 | #define SvSIOK(sv) ((SvFLAGS(sv) & (SVf_IOK|SVf_IVisUV)) == SVf_IOK) | |
693 | #define SvNSIV(sv) ( SvNOK(sv) ? SvNVX(sv) : ( SvSIOK(sv) ? SvIVX(sv) : sv_2nv(sv) ) ) | |
694 | ||
84d4ea48 JH |
695 | PP(pp_sort) |
696 | { | |
20b7effb | 697 | dSP; dMARK; dORIGMARK; |
eb578fdb | 698 | SV **p1 = ORIGMARK+1, **p2; |
c70927a6 | 699 | SSize_t max, i; |
7d49f689 | 700 | AV* av = NULL; |
84d4ea48 | 701 | GV *gv; |
cbbf8932 | 702 | CV *cv = NULL; |
1c23e2bd | 703 | U8 gimme = GIMME_V; |
0bcc34c2 | 704 | OP* const nextop = PL_op->op_next; |
84d4ea48 JH |
705 | I32 overloading = 0; |
706 | bool hasargs = FALSE; | |
2b66f6d3 | 707 | bool copytmps; |
84d4ea48 | 708 | I32 is_xsub = 0; |
901017d6 AL |
709 | const U8 priv = PL_op->op_private; |
710 | const U8 flags = PL_op->op_flags; | |
7b9ef140 | 711 | U32 sort_flags = 0; |
044d25c7 | 712 | I32 all_SIVs = 1, descending = 0; |
84d4ea48 | 713 | |
7b9ef140 | 714 | if ((priv & OPpSORT_DESCEND) != 0) |
044d25c7 | 715 | descending = 1; |
7b9ef140 | 716 | if ((priv & OPpSORT_STABLE) != 0) |
7ea738a9 | 717 | sort_flags |= SORTf_STABLE; |
afe59f35 | 718 | if ((priv & OPpSORT_UNSTABLE) != 0) |
7ea738a9 | 719 | sort_flags |= SORTf_UNSTABLE; |
7b9ef140 | 720 | |
84d4ea48 | 721 | if (gimme != G_ARRAY) { |
7ea738a9 TK |
722 | SP = MARK; |
723 | EXTEND(SP,1); | |
724 | RETPUSHUNDEF; | |
84d4ea48 JH |
725 | } |
726 | ||
727 | ENTER; | |
728 | SAVEVPTR(PL_sortcop); | |
471178c0 | 729 | if (flags & OPf_STACKED) { |
7ea738a9 | 730 | if (flags & OPf_SPECIAL) { |
e6dae479 | 731 | OP *nullop = OpSIBLING(cLISTOP->op_first); /* pass pushmark */ |
932bca29 | 732 | assert(nullop->op_type == OP_NULL); |
7ea738a9 TK |
733 | PL_sortcop = nullop->op_next; |
734 | } | |
735 | else { | |
736 | GV *autogv = NULL; | |
737 | HV *stash; | |
738 | cv = sv_2cv(*++MARK, &stash, &gv, GV_ADD); | |
739 | check_cv: | |
740 | if (cv && SvPOK(cv)) { | |
741 | const char * const proto = SvPV_nolen_const(MUTABLE_SV(cv)); | |
742 | if (proto && strEQ(proto, "$$")) { | |
743 | hasargs = TRUE; | |
744 | } | |
745 | } | |
746 | if (cv && CvISXSUB(cv) && CvXSUB(cv)) { | |
747 | is_xsub = 1; | |
748 | } | |
749 | else if (!(cv && CvROOT(cv))) { | |
750 | if (gv) { | |
751 | goto autoload; | |
752 | } | |
753 | else if (!CvANON(cv) && (gv = CvGV(cv))) { | |
754 | if (cv != GvCV(gv)) cv = GvCV(gv); | |
755 | autoload: | |
756 | if (!autogv && ( | |
757 | autogv = gv_autoload_pvn( | |
758 | GvSTASH(gv), GvNAME(gv), GvNAMELEN(gv), | |
759 | GvNAMEUTF8(gv) ? SVf_UTF8 : 0 | |
760 | ) | |
761 | )) { | |
762 | cv = GvCVu(autogv); | |
763 | goto check_cv; | |
764 | } | |
765 | else { | |
766 | SV *tmpstr = sv_newmortal(); | |
767 | gv_efullname3(tmpstr, gv, NULL); | |
768 | DIE(aTHX_ "Undefined sort subroutine \"%" SVf "\" called", | |
769 | SVfARG(tmpstr)); | |
770 | } | |
771 | } | |
772 | else { | |
773 | DIE(aTHX_ "Undefined subroutine in sort"); | |
774 | } | |
775 | } | |
776 | ||
777 | if (is_xsub) | |
778 | PL_sortcop = (OP*)cv; | |
779 | else | |
780 | PL_sortcop = CvSTART(cv); | |
781 | } | |
84d4ea48 JH |
782 | } |
783 | else { | |
7ea738a9 | 784 | PL_sortcop = NULL; |
84d4ea48 JH |
785 | } |
786 | ||
84721d61 DM |
787 | /* optimiser converts "@a = sort @a" to "sort \@a". In this case, |
788 | * push (@a) onto stack, then assign result back to @a at the end of | |
789 | * this function */ | |
0723351e | 790 | if (priv & OPpSORT_INPLACE) { |
7ea738a9 TK |
791 | assert( MARK+1 == SP && *SP && SvTYPE(*SP) == SVt_PVAV); |
792 | (void)POPMARK; /* remove mark associated with ex-OP_AASSIGN */ | |
793 | av = MUTABLE_AV((*SP)); | |
84721d61 DM |
794 | if (SvREADONLY(av)) |
795 | Perl_croak_no_modify(); | |
7ea738a9 | 796 | max = AvFILL(av) + 1; |
84721d61 | 797 | MEXTEND(SP, max); |
7ea738a9 TK |
798 | if (SvMAGICAL(av)) { |
799 | for (i=0; i < max; i++) { | |
800 | SV **svp = av_fetch(av, i, FALSE); | |
801 | *SP++ = (svp) ? *svp : NULL; | |
802 | } | |
803 | } | |
84721d61 DM |
804 | else { |
805 | SV **svp = AvARRAY(av); | |
806 | assert(svp || max == 0); | |
7ea738a9 | 807 | for (i = 0; i < max; i++) |
84721d61 | 808 | *SP++ = *svp++; |
7ea738a9 | 809 | } |
84721d61 DM |
810 | SP--; |
811 | p1 = p2 = SP - (max-1); | |
fe1bc4cf DM |
812 | } |
813 | else { | |
7ea738a9 TK |
814 | p2 = MARK+1; |
815 | max = SP - MARK; | |
816 | } | |
fe1bc4cf | 817 | |
83a44efe SF |
818 | /* shuffle stack down, removing optional initial cv (p1!=p2), plus |
819 | * any nulls; also stringify or converting to integer or number as | |
820 | * required any args */ | |
ff859a7f | 821 | copytmps = cBOOL(PL_sortcop); |
fe1bc4cf | 822 | for (i=max; i > 0 ; i--) { |
7ea738a9 TK |
823 | if ((*p1 = *p2++)) { /* Weed out nulls. */ |
824 | if (copytmps && SvPADTMP(*p1)) { | |
825 | *p1 = sv_mortalcopy(*p1); | |
826 | } | |
827 | SvTEMP_off(*p1); | |
828 | if (!PL_sortcop) { | |
829 | if (priv & OPpSORT_NUMERIC) { | |
830 | if (priv & OPpSORT_INTEGER) { | |
831 | if (!SvIOK(*p1)) | |
832 | (void)sv_2iv_flags(*p1, SV_GMAGIC|SV_SKIP_OVERLOAD); | |
833 | } | |
834 | else { | |
835 | if (!SvNSIOK(*p1)) | |
836 | (void)sv_2nv_flags(*p1, SV_GMAGIC|SV_SKIP_OVERLOAD); | |
837 | if (all_SIVs && !SvSIOK(*p1)) | |
838 | all_SIVs = 0; | |
839 | } | |
840 | } | |
841 | else { | |
842 | if (!SvPOK(*p1)) | |
843 | (void)sv_2pv_flags(*p1, 0, | |
844 | SV_GMAGIC|SV_CONST_RETURN|SV_SKIP_OVERLOAD); | |
845 | } | |
846 | if (SvAMAGIC(*p1)) | |
847 | overloading = 1; | |
60779a30 | 848 | } |
7ea738a9 TK |
849 | p1++; |
850 | } | |
851 | else | |
852 | max--; | |
84d4ea48 | 853 | } |
fe1bc4cf | 854 | if (max > 1) { |
7ea738a9 TK |
855 | SV **start; |
856 | if (PL_sortcop) { | |
857 | PERL_CONTEXT *cx; | |
858 | const bool oldcatch = CATCH_GET; | |
8ae997c5 | 859 | I32 old_savestack_ix = PL_savestack_ix; |
84d4ea48 | 860 | |
7ea738a9 TK |
861 | SAVEOP(); |
862 | ||
863 | CATCH_SET(TRUE); | |
864 | PUSHSTACKi(PERLSI_SORT); | |
865 | if (!hasargs && !is_xsub) { | |
866 | SAVEGENERICSV(PL_firstgv); | |
867 | SAVEGENERICSV(PL_secondgv); | |
868 | PL_firstgv = MUTABLE_GV(SvREFCNT_inc( | |
869 | gv_fetchpvs("a", GV_ADD|GV_NOTQUAL, SVt_PV) | |
870 | )); | |
871 | PL_secondgv = MUTABLE_GV(SvREFCNT_inc( | |
872 | gv_fetchpvs("b", GV_ADD|GV_NOTQUAL, SVt_PV) | |
873 | )); | |
dc9ef998 TC |
874 | /* make sure the GP isn't removed out from under us for |
875 | * the SAVESPTR() */ | |
876 | save_gp(PL_firstgv, 0); | |
877 | save_gp(PL_secondgv, 0); | |
878 | /* we don't want modifications localized */ | |
879 | GvINTRO_off(PL_firstgv); | |
880 | GvINTRO_off(PL_secondgv); | |
7ea738a9 TK |
881 | SAVEGENERICSV(GvSV(PL_firstgv)); |
882 | SvREFCNT_inc(GvSV(PL_firstgv)); | |
883 | SAVEGENERICSV(GvSV(PL_secondgv)); | |
884 | SvREFCNT_inc(GvSV(PL_secondgv)); | |
885 | } | |
84d4ea48 | 886 | |
33411212 | 887 | gimme = G_SCALAR; |
7ea738a9 TK |
888 | cx = cx_pushblock(CXt_NULL, gimme, PL_stack_base, old_savestack_ix); |
889 | if (!(flags & OPf_SPECIAL)) { | |
890 | cx->cx_type = CXt_SUB|CXp_MULTICALL; | |
891 | cx_pushsub(cx, cv, NULL, hasargs); | |
892 | if (!is_xsub) { | |
893 | PADLIST * const padlist = CvPADLIST(cv); | |
894 | ||
895 | if (++CvDEPTH(cv) >= 2) | |
896 | pad_push(padlist, CvDEPTH(cv)); | |
897 | PAD_SET_CUR_NOSAVE(padlist, CvDEPTH(cv)); | |
898 | ||
899 | if (hasargs) { | |
900 | /* This is mostly copied from pp_entersub */ | |
901 | AV * const av = MUTABLE_AV(PAD_SVl(0)); | |
902 | ||
903 | cx->blk_sub.savearray = GvAV(PL_defgv); | |
904 | GvAV(PL_defgv) = MUTABLE_AV(SvREFCNT_inc_simple(av)); | |
905 | } | |
906 | ||
907 | } | |
908 | } | |
486430a5 | 909 | |
7ea738a9 | 910 | start = p1 - max; |
3edfb5c3 | 911 | Perl_sortsv_flags(aTHX_ start, max, |
7ea738a9 TK |
912 | (is_xsub ? S_sortcv_xsub : hasargs ? S_sortcv_stacked : S_sortcv), |
913 | sort_flags); | |
84d4ea48 | 914 | |
4df352a8 | 915 | /* Reset cx, in case the context stack has been reallocated. */ |
4ebe6e95 | 916 | cx = CX_CUR(); |
4df352a8 | 917 | |
7ea738a9 | 918 | PL_stack_sp = PL_stack_base + cx->blk_oldsp; |
4df352a8 | 919 | |
2f450c1b | 920 | CX_LEAVE_SCOPE(cx); |
7ea738a9 | 921 | if (!(flags & OPf_SPECIAL)) { |
4df352a8 | 922 | assert(CxTYPE(cx) == CXt_SUB); |
a73d8813 | 923 | cx_popsub(cx); |
7ea738a9 | 924 | } |
2f450c1b | 925 | else |
4df352a8 | 926 | assert(CxTYPE(cx) == CXt_NULL); |
2f450c1b | 927 | /* there isn't a POPNULL ! */ |
1dfbe6b4 | 928 | |
7ea738a9 | 929 | cx_popblock(cx); |
5da525e9 | 930 | CX_POP(cx); |
7ea738a9 TK |
931 | POPSTACK; |
932 | CATCH_SET(oldcatch); | |
933 | } | |
934 | else { | |
935 | MEXTEND(SP, 20); /* Can't afford stack realloc on signal. */ | |
936 | start = ORIGMARK+1; | |
433b3e2b TK |
937 | if (priv & OPpSORT_NUMERIC) { |
938 | if ((priv & OPpSORT_INTEGER) || all_SIVs) { | |
939 | if (overloading) | |
044d25c7 TK |
940 | if (descending) |
941 | sortsv_amagic_i_ncmp_desc(aTHX_ start, max, sort_flags); | |
942 | else | |
943 | sortsv_amagic_i_ncmp(aTHX_ start, max, sort_flags); | |
433b3e2b | 944 | else |
044d25c7 TK |
945 | if (descending) |
946 | sortsv_i_ncmp_desc(aTHX_ start, max, sort_flags); | |
947 | else | |
948 | sortsv_i_ncmp(aTHX_ start, max, sort_flags); | |
433b3e2b TK |
949 | } |
950 | else { | |
951 | if (overloading) | |
044d25c7 TK |
952 | if (descending) |
953 | sortsv_amagic_ncmp_desc(aTHX_ start, max, sort_flags); | |
954 | else | |
955 | sortsv_amagic_ncmp(aTHX_ start, max, sort_flags); | |
433b3e2b | 956 | else |
044d25c7 TK |
957 | if (descending) |
958 | sortsv_ncmp_desc(aTHX_ start, max, sort_flags); | |
959 | else | |
960 | sortsv_ncmp(aTHX_ start, max, sort_flags); | |
433b3e2b TK |
961 | } |
962 | } | |
130c5df3 | 963 | #ifdef USE_LOCALE_COLLATE |
433b3e2b TK |
964 | else if(IN_LC_RUNTIME(LC_COLLATE)) { |
965 | if (overloading) | |
044d25c7 TK |
966 | if (descending) |
967 | sortsv_amagic_cmp_locale_desc(aTHX_ start, max, sort_flags); | |
968 | else | |
969 | sortsv_amagic_cmp_locale(aTHX_ start, max, sort_flags); | |
433b3e2b | 970 | else |
044d25c7 TK |
971 | if (descending) |
972 | sortsv_cmp_locale_desc(aTHX_ start, max, sort_flags); | |
973 | else | |
974 | sortsv_cmp_locale(aTHX_ start, max, sort_flags); | |
433b3e2b | 975 | } |
130c5df3 | 976 | #endif |
433b3e2b TK |
977 | else { |
978 | if (overloading) | |
044d25c7 TK |
979 | if (descending) |
980 | sortsv_amagic_cmp_desc(aTHX_ start, max, sort_flags); | |
981 | else | |
982 | sortsv_amagic_cmp(aTHX_ start, max, sort_flags); | |
433b3e2b | 983 | else |
044d25c7 TK |
984 | if (descending) |
985 | sortsv_cmp_desc(aTHX_ start, max, sort_flags); | |
986 | else | |
987 | sortsv_cmp(aTHX_ start, max, sort_flags); | |
433b3e2b | 988 | } |
7ea738a9 TK |
989 | } |
990 | if ((priv & OPpSORT_REVERSE) != 0) { | |
991 | SV **q = start+max-1; | |
992 | while (start < q) { | |
993 | SV * const tmp = *start; | |
994 | *start++ = *q; | |
995 | *q-- = tmp; | |
996 | } | |
997 | } | |
84d4ea48 | 998 | } |
84721d61 DM |
999 | |
1000 | if (av) { | |
1001 | /* copy back result to the array */ | |
1002 | SV** const base = MARK+1; | |
99c9ca9e | 1003 | SSize_t max_minus_one = max - 1; /* attempt to work around mingw bug */ |
84721d61 | 1004 | if (SvMAGICAL(av)) { |
99c9ca9e | 1005 | for (i = 0; i <= max_minus_one; i++) |
84721d61 DM |
1006 | base[i] = newSVsv(base[i]); |
1007 | av_clear(av); | |
99c9ca9e YO |
1008 | if (max_minus_one >= 0) |
1009 | av_extend(av, max_minus_one); | |
1010 | for (i=0; i <= max_minus_one; i++) { | |
84721d61 DM |
1011 | SV * const sv = base[i]; |
1012 | SV ** const didstore = av_store(av, i, sv); | |
1013 | if (SvSMAGICAL(sv)) | |
1014 | mg_set(sv); | |
1015 | if (!didstore) | |
1016 | sv_2mortal(sv); | |
1017 | } | |
1018 | } | |
1019 | else { | |
1020 | /* the elements of av are likely to be the same as the | |
1021 | * (non-refcounted) elements on the stack, just in a different | |
1022 | * order. However, its possible that someone's messed with av | |
1023 | * in the meantime. So bump and unbump the relevant refcounts | |
1024 | * first. | |
1025 | */ | |
99c9ca9e | 1026 | for (i = 0; i <= max_minus_one; i++) { |
45c198c1 DM |
1027 | SV *sv = base[i]; |
1028 | assert(sv); | |
1029 | if (SvREFCNT(sv) > 1) | |
1030 | base[i] = newSVsv(sv); | |
1031 | else | |
1032 | SvREFCNT_inc_simple_void_NN(sv); | |
1033 | } | |
84721d61 | 1034 | av_clear(av); |
99c9ca9e YO |
1035 | if (max_minus_one >= 0) { |
1036 | av_extend(av, max_minus_one); | |
84721d61 DM |
1037 | Copy(base, AvARRAY(av), max, SV*); |
1038 | } | |
99c9ca9e | 1039 | AvFILLp(av) = max_minus_one; |
84721d61 DM |
1040 | AvREIFY_off(av); |
1041 | AvREAL_on(av); | |
1042 | } | |
fe1bc4cf | 1043 | } |
84d4ea48 | 1044 | LEAVE; |
84721d61 | 1045 | PL_stack_sp = ORIGMARK + max; |
84d4ea48 JH |
1046 | return nextop; |
1047 | } | |
1048 | ||
1049 | static I32 | |
31e9e0a3 | 1050 | S_sortcv(pTHX_ SV *const a, SV *const b) |
84d4ea48 | 1051 | { |
901017d6 | 1052 | const I32 oldsaveix = PL_savestack_ix; |
84d4ea48 | 1053 | I32 result; |
ad021bfb | 1054 | PMOP * const pm = PL_curpm; |
a9ea019a | 1055 | COP * const cop = PL_curcop; |
16ada235 | 1056 | SV *olda, *oldb; |
7918f24d NC |
1057 | |
1058 | PERL_ARGS_ASSERT_SORTCV; | |
1059 | ||
16ada235 Z |
1060 | olda = GvSV(PL_firstgv); |
1061 | GvSV(PL_firstgv) = SvREFCNT_inc_simple_NN(a); | |
1062 | SvREFCNT_dec(olda); | |
1063 | oldb = GvSV(PL_secondgv); | |
1064 | GvSV(PL_secondgv) = SvREFCNT_inc_simple_NN(b); | |
1065 | SvREFCNT_dec(oldb); | |
84d4ea48 JH |
1066 | PL_stack_sp = PL_stack_base; |
1067 | PL_op = PL_sortcop; | |
1068 | CALLRUNOPS(aTHX); | |
a9ea019a | 1069 | PL_curcop = cop; |
33411212 DM |
1070 | /* entry zero of a stack is always PL_sv_undef, which |
1071 | * simplifies converting a '()' return into undef in scalar context */ | |
1072 | assert(PL_stack_sp > PL_stack_base || *PL_stack_base == &PL_sv_undef); | |
1073 | result = SvIV(*PL_stack_sp); | |
626ed49c | 1074 | |
53d3542d | 1075 | LEAVE_SCOPE(oldsaveix); |
ad021bfb | 1076 | PL_curpm = pm; |
84d4ea48 JH |
1077 | return result; |
1078 | } | |
1079 | ||
1080 | static I32 | |
31e9e0a3 | 1081 | S_sortcv_stacked(pTHX_ SV *const a, SV *const b) |
84d4ea48 | 1082 | { |
901017d6 | 1083 | const I32 oldsaveix = PL_savestack_ix; |
84d4ea48 | 1084 | I32 result; |
901017d6 | 1085 | AV * const av = GvAV(PL_defgv); |
ad021bfb | 1086 | PMOP * const pm = PL_curpm; |
a9ea019a | 1087 | COP * const cop = PL_curcop; |
84d4ea48 | 1088 | |
7918f24d NC |
1089 | PERL_ARGS_ASSERT_SORTCV_STACKED; |
1090 | ||
8f443ca6 | 1091 | if (AvREAL(av)) { |
7ea738a9 TK |
1092 | av_clear(av); |
1093 | AvREAL_off(av); | |
1094 | AvREIFY_on(av); | |
8f443ca6 | 1095 | } |
84d4ea48 | 1096 | if (AvMAX(av) < 1) { |
7ea738a9 TK |
1097 | SV **ary = AvALLOC(av); |
1098 | if (AvARRAY(av) != ary) { | |
1099 | AvMAX(av) += AvARRAY(av) - AvALLOC(av); | |
1100 | AvARRAY(av) = ary; | |
1101 | } | |
1102 | if (AvMAX(av) < 1) { | |
1103 | Renew(ary,2,SV*); | |
1104 | AvMAX(av) = 1; | |
1105 | AvARRAY(av) = ary; | |
1106 | AvALLOC(av) = ary; | |
1107 | } | |
84d4ea48 JH |
1108 | } |
1109 | AvFILLp(av) = 1; | |
1110 | ||
1111 | AvARRAY(av)[0] = a; | |
1112 | AvARRAY(av)[1] = b; | |
1113 | PL_stack_sp = PL_stack_base; | |
1114 | PL_op = PL_sortcop; | |
1115 | CALLRUNOPS(aTHX); | |
a9ea019a | 1116 | PL_curcop = cop; |
33411212 DM |
1117 | /* entry zero of a stack is always PL_sv_undef, which |
1118 | * simplifies converting a '()' return into undef in scalar context */ | |
1119 | assert(PL_stack_sp > PL_stack_base || *PL_stack_base == &PL_sv_undef); | |
1120 | result = SvIV(*PL_stack_sp); | |
626ed49c | 1121 | |
53d3542d | 1122 | LEAVE_SCOPE(oldsaveix); |
ad021bfb | 1123 | PL_curpm = pm; |
84d4ea48 JH |
1124 | return result; |
1125 | } | |
1126 | ||
1127 | static I32 | |
31e9e0a3 | 1128 | S_sortcv_xsub(pTHX_ SV *const a, SV *const b) |
84d4ea48 | 1129 | { |
20b7effb | 1130 | dSP; |
901017d6 | 1131 | const I32 oldsaveix = PL_savestack_ix; |
ea726b52 | 1132 | CV * const cv=MUTABLE_CV(PL_sortcop); |
84d4ea48 | 1133 | I32 result; |
ad021bfb | 1134 | PMOP * const pm = PL_curpm; |
84d4ea48 | 1135 | |
7918f24d NC |
1136 | PERL_ARGS_ASSERT_SORTCV_XSUB; |
1137 | ||
84d4ea48 JH |
1138 | SP = PL_stack_base; |
1139 | PUSHMARK(SP); | |
1140 | EXTEND(SP, 2); | |
1141 | *++SP = a; | |
1142 | *++SP = b; | |
1143 | PUTBACK; | |
1144 | (void)(*CvXSUB(cv))(aTHX_ cv); | |
33411212 DM |
1145 | /* entry zero of a stack is always PL_sv_undef, which |
1146 | * simplifies converting a '()' return into undef in scalar context */ | |
1147 | assert(PL_stack_sp > PL_stack_base || *PL_stack_base == &PL_sv_undef); | |
84d4ea48 | 1148 | result = SvIV(*PL_stack_sp); |
33411212 | 1149 | |
53d3542d | 1150 | LEAVE_SCOPE(oldsaveix); |
ad021bfb | 1151 | PL_curpm = pm; |
84d4ea48 JH |
1152 | return result; |
1153 | } | |
1154 | ||
1155 | ||
044d25c7 | 1156 | PERL_STATIC_FORCE_INLINE I32 |
31e9e0a3 | 1157 | S_sv_ncmp(pTHX_ SV *const a, SV *const b) |
84d4ea48 | 1158 | { |
427fbfe8 | 1159 | I32 cmp = do_ncmp(a, b); |
7918f24d NC |
1160 | |
1161 | PERL_ARGS_ASSERT_SV_NCMP; | |
1162 | ||
427fbfe8 | 1163 | if (cmp == 2) { |
7ea738a9 TK |
1164 | if (ckWARN(WARN_UNINITIALIZED)) report_uninit(NULL); |
1165 | return 0; | |
f3dab52a | 1166 | } |
427fbfe8 TC |
1167 | |
1168 | return cmp; | |
84d4ea48 JH |
1169 | } |
1170 | ||
044d25c7 TK |
1171 | PERL_STATIC_FORCE_INLINE I32 |
1172 | S_sv_ncmp_desc(pTHX_ SV *const a, SV *const b) | |
1173 | { | |
1174 | PERL_ARGS_ASSERT_SV_NCMP_DESC; | |
1175 | ||
1176 | return -S_sv_ncmp(aTHX_ a, b); | |
1177 | } | |
1178 | ||
1179 | PERL_STATIC_FORCE_INLINE I32 | |
31e9e0a3 | 1180 | S_sv_i_ncmp(pTHX_ SV *const a, SV *const b) |
84d4ea48 | 1181 | { |
901017d6 AL |
1182 | const IV iv1 = SvIV(a); |
1183 | const IV iv2 = SvIV(b); | |
7918f24d NC |
1184 | |
1185 | PERL_ARGS_ASSERT_SV_I_NCMP; | |
1186 | ||
84d4ea48 JH |
1187 | return iv1 < iv2 ? -1 : iv1 > iv2 ? 1 : 0; |
1188 | } | |
901017d6 | 1189 | |
044d25c7 TK |
1190 | PERL_STATIC_FORCE_INLINE I32 |
1191 | S_sv_i_ncmp_desc(pTHX_ SV *const a, SV *const b) | |
1192 | { | |
1193 | PERL_ARGS_ASSERT_SV_I_NCMP_DESC; | |
1194 | ||
1195 | return -S_sv_i_ncmp(aTHX_ a, b); | |
1196 | } | |
1197 | ||
901017d6 | 1198 | #define tryCALL_AMAGICbin(left,right,meth) \ |
79a8d529 | 1199 | (SvAMAGIC(left)||SvAMAGIC(right)) \ |
7ea738a9 TK |
1200 | ? amagic_call(left, right, meth, 0) \ |
1201 | : NULL; | |
84d4ea48 | 1202 | |
659c4b96 | 1203 | #define SORT_NORMAL_RETURN_VALUE(val) (((val) > 0) ? 1 : ((val) ? -1 : 0)) |
eeb9de02 | 1204 | |
044d25c7 | 1205 | PERL_STATIC_FORCE_INLINE I32 |
5aaab254 | 1206 | S_amagic_ncmp(pTHX_ SV *const a, SV *const b) |
84d4ea48 | 1207 | { |
31d632c3 | 1208 | SV * const tmpsv = tryCALL_AMAGICbin(a,b,ncmp_amg); |
7918f24d NC |
1209 | |
1210 | PERL_ARGS_ASSERT_AMAGIC_NCMP; | |
1211 | ||
84d4ea48 | 1212 | if (tmpsv) { |
84d4ea48 | 1213 | if (SvIOK(tmpsv)) { |
901017d6 | 1214 | const I32 i = SvIVX(tmpsv); |
eeb9de02 | 1215 | return SORT_NORMAL_RETURN_VALUE(i); |
84d4ea48 | 1216 | } |
7ea738a9 TK |
1217 | else { |
1218 | const NV d = SvNV(tmpsv); | |
1219 | return SORT_NORMAL_RETURN_VALUE(d); | |
1220 | } | |
84d4ea48 | 1221 | } |
f0f5dc9d | 1222 | return S_sv_ncmp(aTHX_ a, b); |
84d4ea48 JH |
1223 | } |
1224 | ||
044d25c7 TK |
1225 | PERL_STATIC_FORCE_INLINE I32 |
1226 | S_amagic_ncmp_desc(pTHX_ SV *const a, SV *const b) | |
1227 | { | |
1228 | PERL_ARGS_ASSERT_AMAGIC_NCMP_DESC; | |
1229 | ||
1230 | return -S_amagic_ncmp(aTHX_ a, b); | |
1231 | } | |
1232 | ||
1233 | PERL_STATIC_FORCE_INLINE I32 | |
5aaab254 | 1234 | S_amagic_i_ncmp(pTHX_ SV *const a, SV *const b) |
84d4ea48 | 1235 | { |
31d632c3 | 1236 | SV * const tmpsv = tryCALL_AMAGICbin(a,b,ncmp_amg); |
7918f24d NC |
1237 | |
1238 | PERL_ARGS_ASSERT_AMAGIC_I_NCMP; | |
1239 | ||
84d4ea48 | 1240 | if (tmpsv) { |
84d4ea48 | 1241 | if (SvIOK(tmpsv)) { |
901017d6 | 1242 | const I32 i = SvIVX(tmpsv); |
eeb9de02 | 1243 | return SORT_NORMAL_RETURN_VALUE(i); |
84d4ea48 | 1244 | } |
7ea738a9 TK |
1245 | else { |
1246 | const NV d = SvNV(tmpsv); | |
1247 | return SORT_NORMAL_RETURN_VALUE(d); | |
1248 | } | |
84d4ea48 | 1249 | } |
f0f5dc9d | 1250 | return S_sv_i_ncmp(aTHX_ a, b); |
84d4ea48 JH |
1251 | } |
1252 | ||
044d25c7 TK |
1253 | PERL_STATIC_FORCE_INLINE I32 |
1254 | S_amagic_i_ncmp_desc(pTHX_ SV *const a, SV *const b) | |
1255 | { | |
1256 | PERL_ARGS_ASSERT_AMAGIC_I_NCMP_DESC; | |
1257 | ||
1258 | return -S_amagic_i_ncmp(aTHX_ a, b); | |
1259 | } | |
1260 | ||
1261 | PERL_STATIC_FORCE_INLINE I32 | |
5aaab254 | 1262 | S_amagic_cmp(pTHX_ SV *const str1, SV *const str2) |
84d4ea48 | 1263 | { |
31d632c3 | 1264 | SV * const tmpsv = tryCALL_AMAGICbin(str1,str2,scmp_amg); |
7918f24d NC |
1265 | |
1266 | PERL_ARGS_ASSERT_AMAGIC_CMP; | |
1267 | ||
84d4ea48 | 1268 | if (tmpsv) { |
84d4ea48 | 1269 | if (SvIOK(tmpsv)) { |
901017d6 | 1270 | const I32 i = SvIVX(tmpsv); |
eeb9de02 | 1271 | return SORT_NORMAL_RETURN_VALUE(i); |
84d4ea48 | 1272 | } |
7ea738a9 TK |
1273 | else { |
1274 | const NV d = SvNV(tmpsv); | |
1275 | return SORT_NORMAL_RETURN_VALUE(d); | |
1276 | } | |
84d4ea48 JH |
1277 | } |
1278 | return sv_cmp(str1, str2); | |
1279 | } | |
1280 | ||
044d25c7 TK |
1281 | PERL_STATIC_FORCE_INLINE I32 |
1282 | S_amagic_cmp_desc(pTHX_ SV *const str1, SV *const str2) | |
1283 | { | |
1284 | PERL_ARGS_ASSERT_AMAGIC_CMP_DESC; | |
1285 | ||
1286 | return -S_amagic_cmp(aTHX_ str1, str2); | |
1287 | } | |
1288 | ||
1289 | PERL_STATIC_FORCE_INLINE I32 | |
1290 | S_cmp_desc(pTHX_ SV *const str1, SV *const str2) | |
1291 | { | |
1292 | PERL_ARGS_ASSERT_CMP_DESC; | |
1293 | ||
1294 | return -sv_cmp(str1, str2); | |
1295 | } | |
1296 | ||
91191cf7 KW |
1297 | #ifdef USE_LOCALE_COLLATE |
1298 | ||
044d25c7 | 1299 | PERL_STATIC_FORCE_INLINE I32 |
5aaab254 | 1300 | S_amagic_cmp_locale(pTHX_ SV *const str1, SV *const str2) |
84d4ea48 | 1301 | { |
31d632c3 | 1302 | SV * const tmpsv = tryCALL_AMAGICbin(str1,str2,scmp_amg); |
7918f24d NC |
1303 | |
1304 | PERL_ARGS_ASSERT_AMAGIC_CMP_LOCALE; | |
1305 | ||
84d4ea48 | 1306 | if (tmpsv) { |
84d4ea48 | 1307 | if (SvIOK(tmpsv)) { |
901017d6 | 1308 | const I32 i = SvIVX(tmpsv); |
eeb9de02 | 1309 | return SORT_NORMAL_RETURN_VALUE(i); |
84d4ea48 | 1310 | } |
7ea738a9 TK |
1311 | else { |
1312 | const NV d = SvNV(tmpsv); | |
1313 | return SORT_NORMAL_RETURN_VALUE(d); | |
1314 | } | |
84d4ea48 JH |
1315 | } |
1316 | return sv_cmp_locale(str1, str2); | |
1317 | } | |
241d1a3b | 1318 | |
044d25c7 TK |
1319 | PERL_STATIC_FORCE_INLINE I32 |
1320 | S_amagic_cmp_locale_desc(pTHX_ SV *const str1, SV *const str2) | |
1321 | { | |
1322 | PERL_ARGS_ASSERT_AMAGIC_CMP_LOCALE_DESC; | |
1323 | ||
1324 | return -S_amagic_cmp_locale(aTHX_ str1, str2); | |
1325 | } | |
1326 | ||
1327 | PERL_STATIC_FORCE_INLINE I32 | |
1328 | S_cmp_locale_desc(pTHX_ SV *const str1, SV *const str2) | |
1329 | { | |
1330 | PERL_ARGS_ASSERT_CMP_LOCALE_DESC; | |
1331 | ||
1332 | return -sv_cmp_locale(str1, str2); | |
1333 | } | |
1334 | ||
91191cf7 KW |
1335 | #endif |
1336 | ||
241d1a3b | 1337 | /* |
14d04a33 | 1338 | * ex: set ts=8 sts=4 sw=4 et: |
37442d52 | 1339 | */ |