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
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.
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...
15 * [p.931 of _The Lord of the Rings_, VI/ii: "The Land of Shadow"]
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.
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
30 #define PERL_IN_PP_SORT_C
33 #define sv_cmp_static Perl_sv_cmp
34 #define sv_cmp_locale_static Perl_sv_cmp_locale
37 #define SMALLSORT (200)
40 /* Flags for qsortsv and mergesortsv */
42 #define SORTf_STABLE 2
43 #define SORTf_UNSTABLE 8
46 * The mergesort implementation is by Peter M. Mcilroy <pmcilroy@lucent.com>.
48 * The original code was written in conjunction with BSD Computer Software
49 * Research Group at University of California, Berkeley.
51 * See also: "Optimistic Sorting and Information Theoretic Complexity"
53 * SODA (Fourth Annual ACM-SIAM Symposium on Discrete Algorithms),
54 * pp 467-474, Austin, Texas, 25-27 January 1993.
56 * The integration to Perl is by John P. Linderman <jpl.jpl@gmail.com>.
58 * The code can be distributed under the same terms as Perl itself.
63 typedef char * aptr; /* pointer for arithmetic on sizes */
64 typedef SV * gptr; /* pointers in our lists */
66 /* Binary merge internal sort, with a few special mods
67 ** for the special perl environment it now finds itself in.
69 ** Things that were once options have been hotwired
70 ** to values suitable for this use. In particular, we'll always
71 ** initialize looking for natural runs, we'll always produce stable
72 ** output, and we'll always do Peter McIlroy's binary merge.
75 /* Pointer types for arithmetic and storage and convenience casts */
77 #define APTR(P) ((aptr)(P))
78 #define GPTP(P) ((gptr *)(P))
79 #define GPPP(P) ((gptr **)(P))
82 /* byte offset from pointer P to (larger) pointer Q */
83 #define BYTEOFF(P, Q) (APTR(Q) - APTR(P))
85 #define PSIZE sizeof(gptr)
87 /* If PSIZE is power of 2, make PSHIFT that power, if that helps */
90 #define PNELEM(P, Q) (BYTEOFF(P,Q) >> (PSHIFT))
91 #define PNBYTE(N) ((N) << (PSHIFT))
92 #define PINDEX(P, N) (GPTP(APTR(P) + PNBYTE(N)))
94 /* Leave optimization to compiler */
95 #define PNELEM(P, Q) (GPTP(Q) - GPTP(P))
96 #define PNBYTE(N) ((N) * (PSIZE))
97 #define PINDEX(P, N) (GPTP(P) + (N))
100 /* Pointer into other corresponding to pointer into this */
101 #define POTHER(P, THIS, OTHER) GPTP(APTR(OTHER) + BYTEOFF(THIS,P))
103 #define FROMTOUPTO(src, dst, lim) do *dst++ = *src++; while(src<lim)
106 /* Runs are identified by a pointer in the auxiliary list.
107 ** The pointer is at the start of the list,
108 ** and it points to the start of the next list.
109 ** NEXT is used as an lvalue, too.
112 #define NEXT(P) (*GPPP(P))
115 /* PTHRESH is the minimum number of pairs with the same sense to justify
116 ** checking for a run and extending it. Note that PTHRESH counts PAIRS,
117 ** not just elements, so PTHRESH == 8 means a run of 16.
122 /* RTHRESH is the number of elements in a run that must compare low
123 ** to the low element from the opposing run before we justify
124 ** doing a binary rampup instead of single stepping.
125 ** In random input, N in a row low should only happen with
126 ** probability 2^(1-N), so we can risk that we are dealing
127 ** with orderly input without paying much when we aren't.
134 ** Overview of algorithm and variables.
135 ** The array of elements at list1 will be organized into runs of length 2,
136 ** or runs of length >= 2 * PTHRESH. We only try to form long runs when
137 ** PTHRESH adjacent pairs compare in the same way, suggesting overall order.
139 ** Unless otherwise specified, pair pointers address the first of two elements.
141 ** b and b+1 are a pair that compare with sense "sense".
142 ** b is the "bottom" of adjacent pairs that might form a longer run.
144 ** p2 parallels b in the list2 array, where runs are defined by
147 ** t represents the "top" of the adjacent pairs that might extend
148 ** the run beginning at b. Usually, t addresses a pair
149 ** that compares with opposite sense from (b,b+1).
150 ** However, it may also address a singleton element at the end of list1,
151 ** or it may be equal to "last", the first element beyond list1.
153 ** r addresses the Nth pair following b. If this would be beyond t,
154 ** we back it off to t. Only when r is less than t do we consider the
155 ** run long enough to consider checking.
157 ** q addresses a pair such that the pairs at b through q already form a run.
158 ** Often, q will equal b, indicating we only are sure of the pair itself.
159 ** However, a search on the previous cycle may have revealed a longer run,
160 ** so q may be greater than b.
162 ** p is used to work back from a candidate r, trying to reach q,
163 ** which would mean b through r would be a run. If we discover such a run,
164 ** we start q at r and try to push it further towards t.
165 ** If b through r is NOT a run, we detect the wrong order at (p-1,p).
166 ** In any event, after the check (if any), we have two main cases.
168 ** 1) Short run. b <= q < p <= r <= t.
169 ** b through q is a run (perhaps trivial)
170 ** q through p are uninteresting pairs
171 ** p through r is a run
173 ** 2) Long run. b < r <= q < t.
174 ** b through q is a run (of length >= 2 * PTHRESH)
176 ** Note that degenerate cases are not only possible, but likely.
177 ** For example, if the pair following b compares with opposite sense,
178 ** then b == q < p == r == t.
183 dynprep(pTHX_ gptr *list1, gptr *list2, size_t nmemb, const SVCOMPARE_t cmp)
186 gptr *b, *p, *q, *t, *p2;
191 last = PINDEX(b, nmemb);
192 sense = (cmp(aTHX_ *b, *(b+1)) > 0);
193 for (p2 = list2; b < last; ) {
194 /* We just started, or just reversed sense.
195 ** Set t at end of pairs with the prevailing sense.
197 for (p = b+2, t = p; ++p < last; t = ++p) {
198 if ((cmp(aTHX_ *t, *p) > 0) != sense) break;
201 /* Having laid out the playing field, look for long runs */
203 p = r = b + (2 * PTHRESH);
204 if (r >= t) p = r = t; /* too short to care about */
206 while (((cmp(aTHX_ *(p-1), *p) > 0) == sense) &&
209 /* b through r is a (long) run.
210 ** Extend it as far as possible.
213 while (((p += 2) < t) &&
214 ((cmp(aTHX_ *(p-1), *p) > 0) == sense)) q = p;
215 r = p = q + 2; /* no simple pairs, no after-run */
218 if (q > b) { /* run of greater than 2 at b */
222 /* pick up singleton, if possible */
225 ((cmp(aTHX_ *(p-1), *p) > 0) == sense))
226 savep = r = p = q = last;
227 p2 = NEXT(p2) = p2 + (p - b); ++runs;
236 while (q < p) { /* simple pairs */
237 p2 = NEXT(p2) = p2 + 2; ++runs;
244 if (((b = p) == t) && ((t+1) == last)) {
245 NEXT(p2) = p2 + 1; ++runs;
256 /* The original merge sort, in use since 5.7, was as fast as, or faster than,
257 * qsort on many platforms, but slower than qsort, conspicuously so,
258 * on others. The most likely explanation was platform-specific
259 * differences in cache sizes and relative speeds.
261 * The quicksort divide-and-conquer algorithm guarantees that, as the
262 * problem is subdivided into smaller and smaller parts, the parts
263 * fit into smaller (and faster) caches. So it doesn't matter how
264 * many levels of cache exist, quicksort will "find" them, and,
265 * as long as smaller is faster, take advantage of them.
267 * By contrast, consider how the original mergesort algorithm worked.
268 * Suppose we have five runs (each typically of length 2 after dynprep).
277 * Adjacent pairs are merged in "grand sweeps" through the input.
278 * This means, on pass 1, the records in runs 1 and 2 aren't revisited until
279 * runs 3 and 4 are merged and the runs from run 5 have been copied.
280 * The only cache that matters is one large enough to hold *all* the input.
281 * On some platforms, this may be many times slower than smaller caches.
283 * The following pseudo-code uses the same basic merge algorithm,
284 * but in a divide-and-conquer way.
286 * # merge $runs runs at offset $offset of list $list1 into $list2.
287 * # all unmerged runs ($runs == 1) originate in list $base.
289 * my ($offset, $runs, $base, $list1, $list2) = @_;
292 * if ($list1 is $base) copy run to $list2
293 * return offset of end of list (or copy)
295 * $off2 = mgsort2($offset, $runs-($runs/2), $base, $list2, $list1)
296 * mgsort2($off2, $runs/2, $base, $list2, $list1)
297 * merge the adjacent runs at $offset of $list1 into $list2
298 * return the offset of the end of the merged runs
301 * mgsort2(0, $runs, $base, $aux, $base);
303 * For our 5 runs, the tree of calls looks like
312 * and the corresponding activity looks like
314 * copy runs 1 and 2 from base to aux
315 * merge runs 1 and 2 from aux to base
316 * (run 3 is where it belongs, no copy needed)
317 * merge runs 12 and 3 from base to aux
318 * (runs 4 and 5 are where they belong, no copy needed)
319 * merge runs 4 and 5 from base to aux
320 * merge runs 123 and 45 from aux to base
322 * Note that we merge runs 1 and 2 immediately after copying them,
323 * while they are still likely to be in fast cache. Similarly,
324 * run 3 is merged with run 12 while it still may be lingering in cache.
325 * This implementation should therefore enjoy much of the cache-friendly
326 * behavior that quicksort does. In addition, it does less copying
327 * than the original mergesort implementation (only runs 1 and 2 are copied)
328 * and the "balancing" of merges is better (merged runs comprise more nearly
329 * equal numbers of original runs).
331 * The actual cache-friendly implementation will use a pseudo-stack
332 * to avoid recursion, and will unroll processing of runs of length 2,
333 * but it is otherwise similar to the recursive implementation.
337 IV offset; /* offset of 1st of 2 runs at this level */
338 IV runs; /* how many runs must be combined into 1 */
339 } off_runs; /* pseudo-stack element */
343 cmp_desc(pTHX_ gptr const a, gptr const b)
345 return -PL_sort_RealCmp(aTHX_ a, b);
349 =for apidoc sortsv_flags
351 In-place sort an array of SV pointers with the given comparison routine,
352 with various SORTf_* flag options.
357 Perl_sortsv_flags(pTHX_ gptr *base, size_t nmemb, SVCOMPARE_t cmp, U32 flags)
361 gptr *f1, *f2, *t, *b, *p;
365 gptr small[SMALLSORT];
367 off_runs stack[60], *stackp;
368 SVCOMPARE_t savecmp = NULL;
370 PERL_ARGS_ASSERT_SORTSV_FLAGS;
371 if (nmemb <= 1) return; /* sorted trivially */
373 if ((flags & SORTf_DESC) != 0) {
374 savecmp = PL_sort_RealCmp; /* Save current comparison routine, if any */
375 PL_sort_RealCmp = cmp; /* Put comparison routine where cmp_desc can find it */
379 if (nmemb <= SMALLSORT) aux = small; /* use stack for aux array */
380 else { Newx(aux,nmemb,gptr); } /* allocate auxiliary array */
383 stackp->runs = dynprep(aTHX_ base, aux, nmemb, cmp);
384 stackp->offset = offset = 0;
385 which[0] = which[2] = base;
388 /* On levels where both runs have be constructed (stackp->runs == 0),
389 * merge them, and note the offset of their end, in case the offset
390 * is needed at the next level up. Hop up a level, and,
391 * as long as stackp->runs is 0, keep merging.
393 IV runs = stackp->runs;
397 list1 = which[iwhich]; /* area where runs are now */
398 list2 = which[++iwhich]; /* area for merged runs */
401 offset = stackp->offset;
402 f1 = p1 = list1 + offset; /* start of first run */
403 p = tp2 = list2 + offset; /* where merged run will go */
404 t = NEXT(p); /* where first run ends */
405 f2 = l1 = POTHER(t, list2, list1); /* ... on the other side */
406 t = NEXT(t); /* where second runs ends */
407 l2 = POTHER(t, list2, list1); /* ... on the other side */
408 offset = PNELEM(list2, t);
409 while (f1 < l1 && f2 < l2) {
410 /* If head 1 is larger than head 2, find ALL the elements
411 ** in list 2 strictly less than head1, write them all,
412 ** then head 1. Then compare the new heads, and repeat,
413 ** until one or both lists are exhausted.
415 ** In all comparisons (after establishing
416 ** which head to merge) the item to merge
417 ** (at pointer q) is the first operand of
418 ** the comparison. When we want to know
419 ** if "q is strictly less than the other",
422 ** because stability demands that we treat equality
423 ** as high when q comes from l2, and as low when
424 ** q was from l1. So we ask the question by doing
425 ** cmp(q, other) <= sense
426 ** and make sense == 0 when equality should look low,
427 ** and -1 when equality should look high.
431 if (cmp(aTHX_ *f1, *f2) <= 0) {
432 q = f2; b = f1; t = l1;
435 q = f1; b = f2; t = l2;
442 ** Leave t at something strictly
443 ** greater than q (or at the end of the list),
444 ** and b at something strictly less than q.
446 for (i = 1, run = 0 ;;) {
447 if ((p = PINDEX(b, i)) >= t) {
449 if (((p = PINDEX(t, -1)) > b) &&
450 (cmp(aTHX_ *q, *p) <= sense))
454 } else if (cmp(aTHX_ *q, *p) <= sense) {
458 if (++run >= RTHRESH) i += i;
462 /* q is known to follow b and must be inserted before t.
463 ** Increment b, so the range of possibilities is [b,t).
464 ** Round binary split down, to favor early appearance.
465 ** Adjust b and t until q belongs just before t.
470 p = PINDEX(b, (PNELEM(b, t) - 1) / 2);
471 if (cmp(aTHX_ *q, *p) <= sense) {
477 /* Copy all the strictly low elements */
480 FROMTOUPTO(f2, tp2, t);
483 FROMTOUPTO(f1, tp2, t);
489 /* Run out remaining list */
491 if (f2 < l2) FROMTOUPTO(f2, tp2, l2);
492 } else FROMTOUPTO(f1, tp2, l1);
493 p1 = NEXT(p1) = POTHER(tp2, list2, list1);
495 if (--level == 0) goto done;
497 t = list1; list1 = list2; list2 = t; /* swap lists */
498 } while ((runs = stackp->runs) == 0);
502 stackp->runs = 0; /* current run will finish level */
503 /* While there are more than 2 runs remaining,
504 * turn them into exactly 2 runs (at the "other" level),
505 * each made up of approximately half the runs.
506 * Stack the second half for later processing,
507 * and set about producing the first half now.
512 stackp->offset = offset;
513 runs -= stackp->runs = runs / 2;
515 /* We must construct a single run from 1 or 2 runs.
516 * All the original runs are in which[0] == base.
517 * The run we construct must end up in which[level&1].
521 /* Constructing a single run from a single run.
522 * If it's where it belongs already, there's nothing to do.
523 * Otherwise, copy it to where it belongs.
524 * A run of 1 is either a singleton at level 0,
525 * or the second half of a split 3. In neither event
526 * is it necessary to set offset. It will be set by the merge
527 * that immediately follows.
529 if (iwhich) { /* Belongs in aux, currently in base */
530 f1 = b = PINDEX(base, offset); /* where list starts */
531 f2 = PINDEX(aux, offset); /* where list goes */
532 t = NEXT(f2); /* where list will end */
533 offset = PNELEM(aux, t); /* offset thereof */
534 t = PINDEX(base, offset); /* where it currently ends */
535 FROMTOUPTO(f1, f2, t); /* copy */
536 NEXT(b) = t; /* set up parallel pointer */
537 } else if (level == 0) goto done; /* single run at level 0 */
539 /* Constructing a single run from two runs.
540 * The merge code at the top will do that.
541 * We need only make sure the two runs are in the "other" array,
542 * so they'll end up in the correct array after the merge.
546 stackp->offset = offset;
547 stackp->runs = 0; /* take care of both runs, trigger merge */
548 if (!iwhich) { /* Merged runs belong in aux, copy 1st */
549 f1 = b = PINDEX(base, offset); /* where first run starts */
550 f2 = PINDEX(aux, offset); /* where it will be copied */
551 t = NEXT(f2); /* where first run will end */
552 offset = PNELEM(aux, t); /* offset thereof */
553 p = PINDEX(base, offset); /* end of first run */
554 t = NEXT(t); /* where second run will end */
555 t = PINDEX(base, PNELEM(aux, t)); /* where it now ends */
556 FROMTOUPTO(f1, f2, t); /* copy both runs */
557 NEXT(b) = p; /* paralleled pointer for 1st */
558 NEXT(p) = t; /* ... and for second */
563 if (aux != small) Safefree(aux); /* free iff allocated */
564 if (savecmp != NULL) {
565 PL_sort_RealCmp = savecmp; /* Restore current comparison routine, if any */
571 * The quicksort implementation was derived from source code contributed
574 * NOTE: this code was derived from Tom Horsley's qsort replacement
575 * and should not be confused with the original code.
578 /* Copyright (C) Tom Horsley, 1997. All rights reserved.
580 Permission granted to distribute under the same terms as perl which are
583 This program is free software; you can redistribute it and/or modify
584 it under the terms of either:
586 a) the GNU General Public License as published by the Free
587 Software Foundation; either version 1, or (at your option) any
590 b) the "Artistic License" which comes with this Kit.
592 Details on the perl license can be found in the perl source code which
593 may be located via the www.perl.com web page.
595 This is the most wonderfulest possible qsort I can come up with (and
596 still be mostly portable) My (limited) tests indicate it consistently
597 does about 20% fewer calls to compare than does the qsort in the Visual
598 C++ library, other vendors may vary.
600 Some of the ideas in here can be found in "Algorithms" by Sedgewick,
601 others I invented myself (or more likely re-invented since they seemed
602 pretty obvious once I watched the algorithm operate for a while).
604 Most of this code was written while watching the Marlins sweep the Giants
605 in the 1997 National League Playoffs - no Braves fans allowed to use this
606 code (just kidding :-).
608 I realize that if I wanted to be true to the perl tradition, the only
609 comment in this file would be something like:
611 ...they shuffled back towards the rear of the line. 'No, not at the
612 rear!' the slave-driver shouted. 'Three files up. And stay there...
614 However, I really needed to violate that tradition just so I could keep
615 track of what happens myself, not to mention some poor fool trying to
616 understand this years from now :-).
619 /* ********************************************************** Configuration */
621 #ifndef QSORT_ORDER_GUESS
622 #define QSORT_ORDER_GUESS 2 /* Select doubling version of the netBSD trick */
625 /* QSORT_MAX_STACK is the largest number of partitions that can be stacked up for
626 future processing - a good max upper bound is log base 2 of memory size
627 (32 on 32 bit machines, 64 on 64 bit machines, etc). In reality can
628 safely be smaller than that since the program is taking up some space and
629 most operating systems only let you grab some subset of contiguous
630 memory (not to mention that you are normally sorting data larger than
631 1 byte element size :-).
633 #ifndef QSORT_MAX_STACK
634 #define QSORT_MAX_STACK 32
637 /* QSORT_BREAK_EVEN is the size of the largest partition we should insertion sort.
638 Anything bigger and we use qsort. If you make this too small, the qsort
639 will probably break (or become less efficient), because it doesn't expect
640 the middle element of a partition to be the same as the right or left -
641 you have been warned).
643 #ifndef QSORT_BREAK_EVEN
644 #define QSORT_BREAK_EVEN 6
647 /* QSORT_PLAY_SAFE is the size of the largest partition we're willing
648 to go quadratic on. We innoculate larger partitions against
649 quadratic behavior by shuffling them before sorting. This is not
650 an absolute guarantee of non-quadratic behavior, but it would take
651 staggeringly bad luck to pick extreme elements as the pivot
652 from randomized data.
654 #ifndef QSORT_PLAY_SAFE
655 #define QSORT_PLAY_SAFE 255
658 /* ************************************************************* Data Types */
660 /* hold left and right index values of a partition waiting to be sorted (the
661 partition includes both left and right - right is NOT one past the end or
664 struct partition_stack_entry {
667 #ifdef QSORT_ORDER_GUESS
668 int qsort_break_even;
672 /* ******************************************************* Shorthand Macros */
674 /* Note that these macros will be used from inside the qsort function where
675 we happen to know that the variable 'elt_size' contains the size of an
676 array element and the variable 'temp' points to enough space to hold a
677 temp element and the variable 'array' points to the array being sorted
678 and 'compare' is the pointer to the compare routine.
680 Also note that there are very many highly architecture specific ways
681 these might be sped up, but this is simply the most generally portable
682 code I could think of.
685 /* Return < 0 == 0 or > 0 as the value of elt1 is < elt2, == elt2, > elt2
687 #define qsort_cmp(elt1, elt2) \
688 ((*compare)(aTHX_ array[elt1], array[elt2]))
690 #ifdef QSORT_ORDER_GUESS
691 #define QSORT_NOTICE_SWAP swapped++;
693 #define QSORT_NOTICE_SWAP
696 /* swaps contents of array elements elt1, elt2.
698 #define qsort_swap(elt1, elt2) \
701 temp = array[elt1]; \
702 array[elt1] = array[elt2]; \
703 array[elt2] = temp; \
706 /* rotate contents of elt1, elt2, elt3 such that elt1 gets elt2, elt2 gets
707 elt3 and elt3 gets elt1.
709 #define qsort_rotate(elt1, elt2, elt3) \
712 temp = array[elt1]; \
713 array[elt1] = array[elt2]; \
714 array[elt2] = array[elt3]; \
715 array[elt3] = temp; \
718 /* ************************************************************ Debug stuff */
725 return; /* good place to set a breakpoint */
728 #define qsort_assert(t) (void)( (t) || (break_here(), 0) )
735 int (*compare)(const void * elt1, const void * elt2),
736 int pc_left, int pc_right, int u_left, int u_right)
740 qsort_assert(pc_left <= pc_right);
741 qsort_assert(u_right < pc_left);
742 qsort_assert(pc_right < u_left);
743 for (i = u_right + 1; i < pc_left; ++i) {
744 qsort_assert(qsort_cmp(i, pc_left) < 0);
746 for (i = pc_left; i < pc_right; ++i) {
747 qsort_assert(qsort_cmp(i, pc_right) == 0);
749 for (i = pc_right + 1; i < u_left; ++i) {
750 qsort_assert(qsort_cmp(pc_right, i) < 0);
754 #define qsort_all_asserts(PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT) \
755 doqsort_all_asserts(array, num_elts, elt_size, compare, \
756 PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT)
760 #define qsort_assert(t) ((void)0)
762 #define qsort_all_asserts(PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT) ((void)0)
767 =head1 Array Manipulation Functions
771 In-place sort an array of SV pointers with the given comparison routine.
773 Currently this always uses mergesort. See C<L</sortsv_flags>> for a more
780 Perl_sortsv(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp)
782 PERL_ARGS_ASSERT_SORTSV;
784 sortsv_flags(array, nmemb, cmp, 0);
787 #define SvNSIOK(sv) ((SvFLAGS(sv) & SVf_NOK) || ((SvFLAGS(sv) & (SVf_IOK|SVf_IVisUV)) == SVf_IOK))
788 #define SvSIOK(sv) ((SvFLAGS(sv) & (SVf_IOK|SVf_IVisUV)) == SVf_IOK)
789 #define SvNSIV(sv) ( SvNOK(sv) ? SvNVX(sv) : ( SvSIOK(sv) ? SvIVX(sv) : sv_2nv(sv) ) )
793 dSP; dMARK; dORIGMARK;
794 SV **p1 = ORIGMARK+1, **p2;
800 OP* const nextop = PL_op->op_next;
802 bool hasargs = FALSE;
805 const U8 priv = PL_op->op_private;
806 const U8 flags = PL_op->op_flags;
808 void (*sortsvp)(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp, U32 flags)
812 if ((priv & OPpSORT_DESCEND) != 0)
813 sort_flags |= SORTf_DESC;
814 if ((priv & OPpSORT_STABLE) != 0)
815 sort_flags |= SORTf_STABLE;
816 if ((priv & OPpSORT_UNSTABLE) != 0)
817 sort_flags |= SORTf_UNSTABLE;
819 if (gimme != G_ARRAY) {
826 SAVEVPTR(PL_sortcop);
827 if (flags & OPf_STACKED) {
828 if (flags & OPf_SPECIAL) {
829 OP *nullop = OpSIBLING(cLISTOP->op_first); /* pass pushmark */
830 assert(nullop->op_type == OP_NULL);
831 PL_sortcop = nullop->op_next;
836 cv = sv_2cv(*++MARK, &stash, &gv, GV_ADD);
838 if (cv && SvPOK(cv)) {
839 const char * const proto = SvPV_nolen_const(MUTABLE_SV(cv));
840 if (proto && strEQ(proto, "$$")) {
844 if (cv && CvISXSUB(cv) && CvXSUB(cv)) {
847 else if (!(cv && CvROOT(cv))) {
851 else if (!CvANON(cv) && (gv = CvGV(cv))) {
852 if (cv != GvCV(gv)) cv = GvCV(gv);
855 autogv = gv_autoload_pvn(
856 GvSTASH(gv), GvNAME(gv), GvNAMELEN(gv),
857 GvNAMEUTF8(gv) ? SVf_UTF8 : 0
864 SV *tmpstr = sv_newmortal();
865 gv_efullname3(tmpstr, gv, NULL);
866 DIE(aTHX_ "Undefined sort subroutine \"%" SVf "\" called",
871 DIE(aTHX_ "Undefined subroutine in sort");
876 PL_sortcop = (OP*)cv;
878 PL_sortcop = CvSTART(cv);
885 /* optimiser converts "@a = sort @a" to "sort \@a". In this case,
886 * push (@a) onto stack, then assign result back to @a at the end of
888 if (priv & OPpSORT_INPLACE) {
889 assert( MARK+1 == SP && *SP && SvTYPE(*SP) == SVt_PVAV);
890 (void)POPMARK; /* remove mark associated with ex-OP_AASSIGN */
891 av = MUTABLE_AV((*SP));
893 Perl_croak_no_modify();
894 max = AvFILL(av) + 1;
897 for (i=0; i < max; i++) {
898 SV **svp = av_fetch(av, i, FALSE);
899 *SP++ = (svp) ? *svp : NULL;
903 SV **svp = AvARRAY(av);
904 assert(svp || max == 0);
905 for (i = 0; i < max; i++)
909 p1 = p2 = SP - (max-1);
916 /* shuffle stack down, removing optional initial cv (p1!=p2), plus
917 * any nulls; also stringify or converting to integer or number as
918 * required any args */
919 copytmps = cBOOL(PL_sortcop);
920 for (i=max; i > 0 ; i--) {
921 if ((*p1 = *p2++)) { /* Weed out nulls. */
922 if (copytmps && SvPADTMP(*p1)) {
923 *p1 = sv_mortalcopy(*p1);
927 if (priv & OPpSORT_NUMERIC) {
928 if (priv & OPpSORT_INTEGER) {
930 (void)sv_2iv_flags(*p1, SV_GMAGIC|SV_SKIP_OVERLOAD);
934 (void)sv_2nv_flags(*p1, SV_GMAGIC|SV_SKIP_OVERLOAD);
935 if (all_SIVs && !SvSIOK(*p1))
941 (void)sv_2pv_flags(*p1, 0,
942 SV_GMAGIC|SV_CONST_RETURN|SV_SKIP_OVERLOAD);
956 const bool oldcatch = CATCH_GET;
957 I32 old_savestack_ix = PL_savestack_ix;
962 PUSHSTACKi(PERLSI_SORT);
963 if (!hasargs && !is_xsub) {
964 SAVEGENERICSV(PL_firstgv);
965 SAVEGENERICSV(PL_secondgv);
966 PL_firstgv = MUTABLE_GV(SvREFCNT_inc(
967 gv_fetchpvs("a", GV_ADD|GV_NOTQUAL, SVt_PV)
969 PL_secondgv = MUTABLE_GV(SvREFCNT_inc(
970 gv_fetchpvs("b", GV_ADD|GV_NOTQUAL, SVt_PV)
972 /* make sure the GP isn't removed out from under us for
974 save_gp(PL_firstgv, 0);
975 save_gp(PL_secondgv, 0);
976 /* we don't want modifications localized */
977 GvINTRO_off(PL_firstgv);
978 GvINTRO_off(PL_secondgv);
979 SAVEGENERICSV(GvSV(PL_firstgv));
980 SvREFCNT_inc(GvSV(PL_firstgv));
981 SAVEGENERICSV(GvSV(PL_secondgv));
982 SvREFCNT_inc(GvSV(PL_secondgv));
986 cx = cx_pushblock(CXt_NULL, gimme, PL_stack_base, old_savestack_ix);
987 if (!(flags & OPf_SPECIAL)) {
988 cx->cx_type = CXt_SUB|CXp_MULTICALL;
989 cx_pushsub(cx, cv, NULL, hasargs);
991 PADLIST * const padlist = CvPADLIST(cv);
993 if (++CvDEPTH(cv) >= 2)
994 pad_push(padlist, CvDEPTH(cv));
995 PAD_SET_CUR_NOSAVE(padlist, CvDEPTH(cv));
998 /* This is mostly copied from pp_entersub */
999 AV * const av = MUTABLE_AV(PAD_SVl(0));
1001 cx->blk_sub.savearray = GvAV(PL_defgv);
1002 GvAV(PL_defgv) = MUTABLE_AV(SvREFCNT_inc_simple(av));
1009 sortsvp(aTHX_ start, max,
1010 (is_xsub ? S_sortcv_xsub : hasargs ? S_sortcv_stacked : S_sortcv),
1013 /* Reset cx, in case the context stack has been reallocated. */
1016 PL_stack_sp = PL_stack_base + cx->blk_oldsp;
1019 if (!(flags & OPf_SPECIAL)) {
1020 assert(CxTYPE(cx) == CXt_SUB);
1024 assert(CxTYPE(cx) == CXt_NULL);
1025 /* there isn't a POPNULL ! */
1030 CATCH_SET(oldcatch);
1033 MEXTEND(SP, 20); /* Can't afford stack realloc on signal. */
1035 sortsvp(aTHX_ start, max,
1036 (priv & OPpSORT_NUMERIC)
1037 ? ( ( ( priv & OPpSORT_INTEGER) || all_SIVs)
1038 ? ( overloading ? S_amagic_i_ncmp : S_sv_i_ncmp)
1039 : ( overloading ? S_amagic_ncmp : S_sv_ncmp ) )
1041 #ifdef USE_LOCALE_COLLATE
1042 IN_LC_RUNTIME(LC_COLLATE)
1044 ? (SVCOMPARE_t)S_amagic_cmp_locale
1045 : (SVCOMPARE_t)sv_cmp_locale_static)
1048 ( overloading ? (SVCOMPARE_t)S_amagic_cmp : (SVCOMPARE_t)sv_cmp_static)),
1051 if ((priv & OPpSORT_REVERSE) != 0) {
1052 SV **q = start+max-1;
1054 SV * const tmp = *start;
1062 /* copy back result to the array */
1063 SV** const base = MARK+1;
1064 if (SvMAGICAL(av)) {
1065 for (i = 0; i < max; i++)
1066 base[i] = newSVsv(base[i]);
1069 for (i=0; i < max; i++) {
1070 SV * const sv = base[i];
1071 SV ** const didstore = av_store(av, i, sv);
1079 /* the elements of av are likely to be the same as the
1080 * (non-refcounted) elements on the stack, just in a different
1081 * order. However, its possible that someone's messed with av
1082 * in the meantime. So bump and unbump the relevant refcounts
1085 for (i = 0; i < max; i++) {
1088 if (SvREFCNT(sv) > 1)
1089 base[i] = newSVsv(sv);
1091 SvREFCNT_inc_simple_void_NN(sv);
1096 Copy(base, AvARRAY(av), max, SV*);
1098 AvFILLp(av) = max - 1;
1104 PL_stack_sp = ORIGMARK + max;
1109 S_sortcv(pTHX_ SV *const a, SV *const b)
1111 const I32 oldsaveix = PL_savestack_ix;
1113 PMOP * const pm = PL_curpm;
1114 COP * const cop = PL_curcop;
1117 PERL_ARGS_ASSERT_SORTCV;
1119 olda = GvSV(PL_firstgv);
1120 GvSV(PL_firstgv) = SvREFCNT_inc_simple_NN(a);
1122 oldb = GvSV(PL_secondgv);
1123 GvSV(PL_secondgv) = SvREFCNT_inc_simple_NN(b);
1125 PL_stack_sp = PL_stack_base;
1129 /* entry zero of a stack is always PL_sv_undef, which
1130 * simplifies converting a '()' return into undef in scalar context */
1131 assert(PL_stack_sp > PL_stack_base || *PL_stack_base == &PL_sv_undef);
1132 result = SvIV(*PL_stack_sp);
1134 LEAVE_SCOPE(oldsaveix);
1140 S_sortcv_stacked(pTHX_ SV *const a, SV *const b)
1142 const I32 oldsaveix = PL_savestack_ix;
1144 AV * const av = GvAV(PL_defgv);
1145 PMOP * const pm = PL_curpm;
1146 COP * const cop = PL_curcop;
1148 PERL_ARGS_ASSERT_SORTCV_STACKED;
1155 if (AvMAX(av) < 1) {
1156 SV **ary = AvALLOC(av);
1157 if (AvARRAY(av) != ary) {
1158 AvMAX(av) += AvARRAY(av) - AvALLOC(av);
1161 if (AvMAX(av) < 1) {
1172 PL_stack_sp = PL_stack_base;
1176 /* entry zero of a stack is always PL_sv_undef, which
1177 * simplifies converting a '()' return into undef in scalar context */
1178 assert(PL_stack_sp > PL_stack_base || *PL_stack_base == &PL_sv_undef);
1179 result = SvIV(*PL_stack_sp);
1181 LEAVE_SCOPE(oldsaveix);
1187 S_sortcv_xsub(pTHX_ SV *const a, SV *const b)
1190 const I32 oldsaveix = PL_savestack_ix;
1191 CV * const cv=MUTABLE_CV(PL_sortcop);
1193 PMOP * const pm = PL_curpm;
1195 PERL_ARGS_ASSERT_SORTCV_XSUB;
1203 (void)(*CvXSUB(cv))(aTHX_ cv);
1204 /* entry zero of a stack is always PL_sv_undef, which
1205 * simplifies converting a '()' return into undef in scalar context */
1206 assert(PL_stack_sp > PL_stack_base || *PL_stack_base == &PL_sv_undef);
1207 result = SvIV(*PL_stack_sp);
1209 LEAVE_SCOPE(oldsaveix);
1216 S_sv_ncmp(pTHX_ SV *const a, SV *const b)
1218 I32 cmp = do_ncmp(a, b);
1220 PERL_ARGS_ASSERT_SV_NCMP;
1223 if (ckWARN(WARN_UNINITIALIZED)) report_uninit(NULL);
1231 S_sv_i_ncmp(pTHX_ SV *const a, SV *const b)
1233 const IV iv1 = SvIV(a);
1234 const IV iv2 = SvIV(b);
1236 PERL_ARGS_ASSERT_SV_I_NCMP;
1238 return iv1 < iv2 ? -1 : iv1 > iv2 ? 1 : 0;
1241 #define tryCALL_AMAGICbin(left,right,meth) \
1242 (SvAMAGIC(left)||SvAMAGIC(right)) \
1243 ? amagic_call(left, right, meth, 0) \
1246 #define SORT_NORMAL_RETURN_VALUE(val) (((val) > 0) ? 1 : ((val) ? -1 : 0))
1249 S_amagic_ncmp(pTHX_ SV *const a, SV *const b)
1251 SV * const tmpsv = tryCALL_AMAGICbin(a,b,ncmp_amg);
1253 PERL_ARGS_ASSERT_AMAGIC_NCMP;
1257 const I32 i = SvIVX(tmpsv);
1258 return SORT_NORMAL_RETURN_VALUE(i);
1261 const NV d = SvNV(tmpsv);
1262 return SORT_NORMAL_RETURN_VALUE(d);
1265 return S_sv_ncmp(aTHX_ a, b);
1269 S_amagic_i_ncmp(pTHX_ SV *const a, SV *const b)
1271 SV * const tmpsv = tryCALL_AMAGICbin(a,b,ncmp_amg);
1273 PERL_ARGS_ASSERT_AMAGIC_I_NCMP;
1277 const I32 i = SvIVX(tmpsv);
1278 return SORT_NORMAL_RETURN_VALUE(i);
1281 const NV d = SvNV(tmpsv);
1282 return SORT_NORMAL_RETURN_VALUE(d);
1285 return S_sv_i_ncmp(aTHX_ a, b);
1289 S_amagic_cmp(pTHX_ SV *const str1, SV *const str2)
1291 SV * const tmpsv = tryCALL_AMAGICbin(str1,str2,scmp_amg);
1293 PERL_ARGS_ASSERT_AMAGIC_CMP;
1297 const I32 i = SvIVX(tmpsv);
1298 return SORT_NORMAL_RETURN_VALUE(i);
1301 const NV d = SvNV(tmpsv);
1302 return SORT_NORMAL_RETURN_VALUE(d);
1305 return sv_cmp(str1, str2);
1308 #ifdef USE_LOCALE_COLLATE
1311 S_amagic_cmp_locale(pTHX_ SV *const str1, SV *const str2)
1313 SV * const tmpsv = tryCALL_AMAGICbin(str1,str2,scmp_amg);
1315 PERL_ARGS_ASSERT_AMAGIC_CMP_LOCALE;
1319 const I32 i = SvIVX(tmpsv);
1320 return SORT_NORMAL_RETURN_VALUE(i);
1323 const NV d = SvNV(tmpsv);
1324 return SORT_NORMAL_RETURN_VALUE(d);
1327 return sv_cmp_locale(str1, str2);
1333 * ex: set ts=8 sts=4 sw=4 et: