xref: /openbsd-src/gnu/usr.bin/perl/pp_sort.c (revision db3296cf5c1dd9058ceecc3a29fe4aaa0bd26000)
1 /*    pp_sort.c
2  *
3  *    Copyright (c) 1991-2002, Larry Wall
4  *
5  *    You may distribute under the terms of either the GNU General Public
6  *    License or the Artistic License, as specified in the README file.
7  *
8  */
9 
10 /*
11  *   ...they shuffled back towards the rear of the line. 'No, not at the
12  *   rear!'  the slave-driver shouted. 'Three files up. And stay there...
13  */
14 
15 #include "EXTERN.h"
16 #define PERL_IN_PP_SORT_C
17 #include "perl.h"
18 
19 #if defined(UNDER_CE)
20 /* looks like 'small' is reserved word for WINCE (or somesuch)*/
21 #define	small xsmall
22 #endif
23 
24 static I32 sortcv(pTHX_ SV *a, SV *b);
25 static I32 sortcv_stacked(pTHX_ SV *a, SV *b);
26 static I32 sortcv_xsub(pTHX_ SV *a, SV *b);
27 static I32 sv_ncmp(pTHX_ SV *a, SV *b);
28 static I32 sv_i_ncmp(pTHX_ SV *a, SV *b);
29 static I32 amagic_ncmp(pTHX_ SV *a, SV *b);
30 static I32 amagic_i_ncmp(pTHX_ SV *a, SV *b);
31 static I32 amagic_cmp(pTHX_ SV *a, SV *b);
32 static I32 amagic_cmp_locale(pTHX_ SV *a, SV *b);
33 
34 #define sv_cmp_static Perl_sv_cmp
35 #define sv_cmp_locale_static Perl_sv_cmp_locale
36 
37 #define SORTHINTS(hintsv) \
38     (((hintsv) = GvSV(gv_fetchpv("sort::hints", GV_ADDMULTI, SVt_IV))), \
39     (SvIOK(hintsv) ? ((I32)SvIV(hintsv)) : 0))
40 
41 #ifndef SMALLSORT
42 #define	SMALLSORT (200)
43 #endif
44 
45 /*
46  * The mergesort implementation is by Peter M. Mcilroy <pmcilroy@lucent.com>.
47  *
48  * The original code was written in conjunction with BSD Computer Software
49  * Research Group at University of California, Berkeley.
50  *
51  * See also: "Optimistic Merge Sort" (SODA '92)
52  *
53  * The integration to Perl is by John P. Linderman <jpl@research.att.com>.
54  *
55  * The code can be distributed under the same terms as Perl itself.
56  *
57  */
58 
59 
60 typedef char * aptr;		/* pointer for arithmetic on sizes */
61 typedef SV * gptr;		/* pointers in our lists */
62 
63 /* Binary merge internal sort, with a few special mods
64 ** for the special perl environment it now finds itself in.
65 **
66 ** Things that were once options have been hotwired
67 ** to values suitable for this use.  In particular, we'll always
68 ** initialize looking for natural runs, we'll always produce stable
69 ** output, and we'll always do Peter McIlroy's binary merge.
70 */
71 
72 /* Pointer types for arithmetic and storage and convenience casts */
73 
74 #define	APTR(P)	((aptr)(P))
75 #define	GPTP(P)	((gptr *)(P))
76 #define GPPP(P) ((gptr **)(P))
77 
78 
79 /* byte offset from pointer P to (larger) pointer Q */
80 #define	BYTEOFF(P, Q) (APTR(Q) - APTR(P))
81 
82 #define PSIZE sizeof(gptr)
83 
84 /* If PSIZE is power of 2, make PSHIFT that power, if that helps */
85 
86 #ifdef	PSHIFT
87 #define	PNELEM(P, Q)	(BYTEOFF(P,Q) >> (PSHIFT))
88 #define	PNBYTE(N)	((N) << (PSHIFT))
89 #define	PINDEX(P, N)	(GPTP(APTR(P) + PNBYTE(N)))
90 #else
91 /* Leave optimization to compiler */
92 #define	PNELEM(P, Q)	(GPTP(Q) - GPTP(P))
93 #define	PNBYTE(N)	((N) * (PSIZE))
94 #define	PINDEX(P, N)	(GPTP(P) + (N))
95 #endif
96 
97 /* Pointer into other corresponding to pointer into this */
98 #define	POTHER(P, THIS, OTHER) GPTP(APTR(OTHER) + BYTEOFF(THIS,P))
99 
100 #define FROMTOUPTO(src, dst, lim) do *dst++ = *src++; while(src<lim)
101 
102 
103 /* Runs are identified by a pointer in the auxilliary list.
104 ** The pointer is at the start of the list,
105 ** and it points to the start of the next list.
106 ** NEXT is used as an lvalue, too.
107 */
108 
109 #define	NEXT(P)		(*GPPP(P))
110 
111 
112 /* PTHRESH is the minimum number of pairs with the same sense to justify
113 ** checking for a run and extending it.  Note that PTHRESH counts PAIRS,
114 ** not just elements, so PTHRESH == 8 means a run of 16.
115 */
116 
117 #define	PTHRESH (8)
118 
119 /* RTHRESH is the number of elements in a run that must compare low
120 ** to the low element from the opposing run before we justify
121 ** doing a binary rampup instead of single stepping.
122 ** In random input, N in a row low should only happen with
123 ** probability 2^(1-N), so we can risk that we are dealing
124 ** with orderly input without paying much when we aren't.
125 */
126 
127 #define RTHRESH (6)
128 
129 
130 /*
131 ** Overview of algorithm and variables.
132 ** The array of elements at list1 will be organized into runs of length 2,
133 ** or runs of length >= 2 * PTHRESH.  We only try to form long runs when
134 ** PTHRESH adjacent pairs compare in the same way, suggesting overall order.
135 **
136 ** Unless otherwise specified, pair pointers address the first of two elements.
137 **
138 ** b and b+1 are a pair that compare with sense ``sense''.
139 ** b is the ``bottom'' of adjacent pairs that might form a longer run.
140 **
141 ** p2 parallels b in the list2 array, where runs are defined by
142 ** a pointer chain.
143 **
144 ** t represents the ``top'' of the adjacent pairs that might extend
145 ** the run beginning at b.  Usually, t addresses a pair
146 ** that compares with opposite sense from (b,b+1).
147 ** However, it may also address a singleton element at the end of list1,
148 ** or it may be equal to ``last'', the first element beyond list1.
149 **
150 ** r addresses the Nth pair following b.  If this would be beyond t,
151 ** we back it off to t.  Only when r is less than t do we consider the
152 ** run long enough to consider checking.
153 **
154 ** q addresses a pair such that the pairs at b through q already form a run.
155 ** Often, q will equal b, indicating we only are sure of the pair itself.
156 ** However, a search on the previous cycle may have revealed a longer run,
157 ** so q may be greater than b.
158 **
159 ** p is used to work back from a candidate r, trying to reach q,
160 ** which would mean b through r would be a run.  If we discover such a run,
161 ** we start q at r and try to push it further towards t.
162 ** If b through r is NOT a run, we detect the wrong order at (p-1,p).
163 ** In any event, after the check (if any), we have two main cases.
164 **
165 ** 1) Short run.  b <= q < p <= r <= t.
166 **	b through q is a run (perhaps trivial)
167 **	q through p are uninteresting pairs
168 **	p through r is a run
169 **
170 ** 2) Long run.  b < r <= q < t.
171 **	b through q is a run (of length >= 2 * PTHRESH)
172 **
173 ** Note that degenerate cases are not only possible, but likely.
174 ** For example, if the pair following b compares with opposite sense,
175 ** then b == q < p == r == t.
176 */
177 
178 
179 static IV
180 dynprep(pTHX_ gptr *list1, gptr *list2, size_t nmemb, SVCOMPARE_t cmp)
181 {
182     I32 sense;
183     register gptr *b, *p, *q, *t, *p2;
184     register gptr c, *last, *r;
185     gptr *savep;
186     IV runs = 0;
187 
188     b = list1;
189     last = PINDEX(b, nmemb);
190     sense = (cmp(aTHX_ *b, *(b+1)) > 0);
191     for (p2 = list2; b < last; ) {
192 	/* We just started, or just reversed sense.
193 	** Set t at end of pairs with the prevailing sense.
194 	*/
195 	for (p = b+2, t = p; ++p < last; t = ++p) {
196 	    if ((cmp(aTHX_ *t, *p) > 0) != sense) break;
197 	}
198 	q = b;
199 	/* Having laid out the playing field, look for long runs */
200 	do {
201 	    p = r = b + (2 * PTHRESH);
202 	    if (r >= t) p = r = t;	/* too short to care about */
203 	    else {
204 		while (((cmp(aTHX_ *(p-1), *p) > 0) == sense) &&
205 		       ((p -= 2) > q));
206 		if (p <= q) {
207 		    /* b through r is a (long) run.
208 		    ** Extend it as far as possible.
209 		    */
210 		    p = q = r;
211 		    while (((p += 2) < t) &&
212 			   ((cmp(aTHX_ *(p-1), *p) > 0) == sense)) q = p;
213 		    r = p = q + 2;	/* no simple pairs, no after-run */
214 		}
215 	    }
216 	    if (q > b) {		/* run of greater than 2 at b */
217 		savep = p;
218 		p = q += 2;
219 		/* pick up singleton, if possible */
220 		if ((p == t) &&
221 		    ((t + 1) == last) &&
222 		    ((cmp(aTHX_ *(p-1), *p) > 0) == sense))
223 		    savep = r = p = q = last;
224 		p2 = NEXT(p2) = p2 + (p - b); ++runs;
225 		if (sense) while (b < --p) {
226 		    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 		    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;
247     }
248     return runs;
249 }
250 
251 
252 /* The original merge sort, in use since 5.7, was as fast as, or faster than,
253  * qsort on many platforms, but slower than qsort, conspicuously so,
254  * on others.  The most likely explanation was platform-specific
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,
261  * as long as smaller is faster, take advanatge of them.
262  *
263  * By contrast, consider how the original mergesort algorithm worked.
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.
330  */
331 
332 typedef struct {
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 */
336 
337 STATIC void
338 S_mergesortsv(pTHX_ gptr *base, size_t nmemb, SVCOMPARE_t cmp)
339 {
340     IV i, run, runs, offset;
341     I32 sense, level;
342     int iwhich;
343     register gptr *f1, *f2, *t, *b, *p, *tp2, *l1, *l2, *q;
344     gptr *aux, *list1, *list2;
345     gptr *p1;
346     gptr small[SMALLSORT];
347     gptr *which[3];
348     off_runs stack[60], *stackp;
349 
350     if (nmemb <= 1) return;			/* sorted trivially */
351     if (nmemb <= SMALLSORT) aux = small;	/* use stack for aux array */
352     else { New(799,aux,nmemb,gptr); }		/* allocate auxilliary array */
353     level = 0;
354     stackp = stack;
355     stackp->runs = dynprep(aTHX_ base, aux, nmemb, cmp);
356     stackp->offset = offset = 0;
357     which[0] = which[2] = base;
358     which[1] = aux;
359     for (;;) {
360 	/* On levels where both runs have be constructed (stackp->runs == 0),
361 	 * merge them, and note the offset of their end, in case the offset
362 	 * is needed at the next level up.  Hop up a level, and,
363 	 * as long as stackp->runs is 0, keep merging.
364 	 */
365 	if ((runs = stackp->runs) == 0) {
366 	    iwhich = level & 1;
367 	    list1 = which[iwhich];		/* area where runs are now */
368 	    list2 = which[++iwhich];		/* area for merged runs */
369 	    do {
370 		offset = stackp->offset;
371 		f1 = p1 = list1 + offset;		/* start of first run */
372 		p = tp2 = list2 + offset;	/* where merged run will go */
373 		t = NEXT(p);			/* where first run ends */
374 		f2 = l1 = POTHER(t, list2, list1); /* ... on the other side */
375 		t = NEXT(t);			/* where second runs ends */
376 		l2 = POTHER(t, list2, list1);	/* ... on the other side */
377 		offset = PNELEM(list2, t);
378 		while (f1 < l1 && f2 < l2) {
379 		    /* If head 1 is larger than head 2, find ALL the elements
380 		    ** in list 2 strictly less than head1, write them all,
381 		    ** then head 1.  Then compare the new heads, and repeat,
382 		    ** until one or both lists are exhausted.
383 		    **
384 		    ** In all comparisons (after establishing
385 		    ** which head to merge) the item to merge
386 		    ** (at pointer q) is the first operand of
387 		    ** the comparison.  When we want to know
388 		    ** if ``q is strictly less than the other'',
389 		    ** we can't just do
390 		    **    cmp(q, other) < 0
391 		    ** because stability demands that we treat equality
392 		    ** as high when q comes from l2, and as low when
393 		    ** q was from l1.  So we ask the question by doing
394 		    **    cmp(q, other) <= sense
395 		    ** and make sense == 0 when equality should look low,
396 		    ** and -1 when equality should look high.
397 		    */
398 
399 
400 		    if (cmp(aTHX_ *f1, *f2) <= 0) {
401 			q = f2; b = f1; t = l1;
402 			sense = -1;
403 		    } else {
404 			q = f1; b = f2; t = l2;
405 			sense = 0;
406 		    }
407 
408 
409 		    /* ramp up
410 		    **
411 		    ** Leave t at something strictly
412 		    ** greater than q (or at the end of the list),
413 		    ** and b at something strictly less than q.
414 		    */
415 		    for (i = 1, run = 0 ;;) {
416 			if ((p = PINDEX(b, i)) >= t) {
417 			    /* off the end */
418 			    if (((p = PINDEX(t, -1)) > b) &&
419 				(cmp(aTHX_ *q, *p) <= sense))
420 				 t = p;
421 			    else b = p;
422 			    break;
423 			} else if (cmp(aTHX_ *q, *p) <= sense) {
424 			    t = p;
425 			    break;
426 			} else b = p;
427 			if (++run >= RTHRESH) i += i;
428 		    }
429 
430 
431 		    /* q is known to follow b and must be inserted before t.
432 		    ** Increment b, so the range of possibilities is [b,t).
433 		    ** Round binary split down, to favor early appearance.
434 		    ** Adjust b and t until q belongs just before t.
435 		    */
436 
437 		    b++;
438 		    while (b < t) {
439 			p = PINDEX(b, (PNELEM(b, t) - 1) / 2);
440 			if (cmp(aTHX_ *q, *p) <= sense) {
441 			    t = p;
442 			} else b = p + 1;
443 		    }
444 
445 
446 		    /* Copy all the strictly low elements */
447 
448 		    if (q == f1) {
449 			FROMTOUPTO(f2, tp2, t);
450 			*tp2++ = *f1++;
451 		    } else {
452 			FROMTOUPTO(f1, tp2, t);
453 			*tp2++ = *f2++;
454 		    }
455 		}
456 
457 
458 		/* Run out remaining list */
459 		if (f1 == l1) {
460 		       if (f2 < l2) FROMTOUPTO(f2, tp2, l2);
461 		} else              FROMTOUPTO(f1, tp2, l1);
462 		p1 = NEXT(p1) = POTHER(tp2, list2, list1);
463 
464 		if (--level == 0) goto done;
465 		--stackp;
466 		t = list1; list1 = list2; list2 = t;	/* swap lists */
467 	    } while ((runs = stackp->runs) == 0);
468 	}
469 
470 
471 	stackp->runs = 0;		/* current run will finish level */
472 	/* While there are more than 2 runs remaining,
473 	 * turn them into exactly 2 runs (at the "other" level),
474 	 * each made up of approximately half the runs.
475 	 * Stack the second half for later processing,
476 	 * and set about producing the first half now.
477 	 */
478 	while (runs > 2) {
479 	    ++level;
480 	    ++stackp;
481 	    stackp->offset = offset;
482 	    runs -= stackp->runs = runs / 2;
483 	}
484 	/* We must construct a single run from 1 or 2 runs.
485 	 * All the original runs are in which[0] == base.
486 	 * The run we construct must end up in which[level&1].
487 	 */
488 	iwhich = level & 1;
489 	if (runs == 1) {
490 	    /* Constructing a single run from a single run.
491 	     * If it's where it belongs already, there's nothing to do.
492 	     * Otherwise, copy it to where it belongs.
493 	     * A run of 1 is either a singleton at level 0,
494 	     * or the second half of a split 3.  In neither event
495 	     * is it necessary to set offset.  It will be set by the merge
496 	     * that immediately follows.
497 	     */
498 	    if (iwhich) {	/* Belongs in aux, currently in base */
499 		f1 = b = PINDEX(base, offset);	/* where list starts */
500 		f2 = PINDEX(aux, offset);	/* where list goes */
501 		t = NEXT(f2);			/* where list will end */
502 		offset = PNELEM(aux, t);	/* offset thereof */
503 		t = PINDEX(base, offset);	/* where it currently ends */
504 		FROMTOUPTO(f1, f2, t);		/* copy */
505 		NEXT(b) = t;			/* set up parallel pointer */
506 	    } else if (level == 0) goto done;	/* single run at level 0 */
507 	} else {
508 	    /* Constructing a single run from two runs.
509 	     * The merge code at the top will do that.
510 	     * We need only make sure the two runs are in the "other" array,
511 	     * so they'll end up in the correct array after the merge.
512 	     */
513 	    ++level;
514 	    ++stackp;
515 	    stackp->offset = offset;
516 	    stackp->runs = 0;	/* take care of both runs, trigger merge */
517 	    if (!iwhich) {	/* Merged runs belong in aux, copy 1st */
518 		f1 = b = PINDEX(base, offset);	/* where first run starts */
519 		f2 = PINDEX(aux, offset);	/* where it will be copied */
520 		t = NEXT(f2);			/* where first run will end */
521 		offset = PNELEM(aux, t);	/* offset thereof */
522 		p = PINDEX(base, offset);	/* end of first run */
523 		t = NEXT(t);			/* where second run will end */
524 		t = PINDEX(base, PNELEM(aux, t)); /* where it now ends */
525 		FROMTOUPTO(f1, f2, t);		/* copy both runs */
526 		NEXT(b) = p;			/* paralled pointer for 1st */
527 		NEXT(p) = t;			/* ... and for second */
528 	    }
529 	}
530     }
531 done:
532     if (aux != small) Safefree(aux);	/* free iff allocated */
533     return;
534 }
535 
536 /*
537  * The quicksort implementation was derived from source code contributed
538  * by Tom Horsley.
539  *
540  * NOTE: this code was derived from Tom Horsley's qsort replacement
541  * and should not be confused with the original code.
542  */
543 
544 /* Copyright (C) Tom Horsley, 1997. All rights reserved.
545 
546    Permission granted to distribute under the same terms as perl which are
547    (briefly):
548 
549     This program is free software; you can redistribute it and/or modify
550     it under the terms of either:
551 
552 	a) the GNU General Public License as published by the Free
553 	Software Foundation; either version 1, or (at your option) any
554 	later version, or
555 
556 	b) the "Artistic License" which comes with this Kit.
557 
558    Details on the perl license can be found in the perl source code which
559    may be located via the www.perl.com web page.
560 
561    This is the most wonderfulest possible qsort I can come up with (and
562    still be mostly portable) My (limited) tests indicate it consistently
563    does about 20% fewer calls to compare than does the qsort in the Visual
564    C++ library, other vendors may vary.
565 
566    Some of the ideas in here can be found in "Algorithms" by Sedgewick,
567    others I invented myself (or more likely re-invented since they seemed
568    pretty obvious once I watched the algorithm operate for a while).
569 
570    Most of this code was written while watching the Marlins sweep the Giants
571    in the 1997 National League Playoffs - no Braves fans allowed to use this
572    code (just kidding :-).
573 
574    I realize that if I wanted to be true to the perl tradition, the only
575    comment in this file would be something like:
576 
577    ...they shuffled back towards the rear of the line. 'No, not at the
578    rear!'  the slave-driver shouted. 'Three files up. And stay there...
579 
580    However, I really needed to violate that tradition just so I could keep
581    track of what happens myself, not to mention some poor fool trying to
582    understand this years from now :-).
583 */
584 
585 /* ********************************************************** Configuration */
586 
587 #ifndef QSORT_ORDER_GUESS
588 #define QSORT_ORDER_GUESS 2	/* Select doubling version of the netBSD trick */
589 #endif
590 
591 /* QSORT_MAX_STACK is the largest number of partitions that can be stacked up for
592    future processing - a good max upper bound is log base 2 of memory size
593    (32 on 32 bit machines, 64 on 64 bit machines, etc). In reality can
594    safely be smaller than that since the program is taking up some space and
595    most operating systems only let you grab some subset of contiguous
596    memory (not to mention that you are normally sorting data larger than
597    1 byte element size :-).
598 */
599 #ifndef QSORT_MAX_STACK
600 #define QSORT_MAX_STACK 32
601 #endif
602 
603 /* QSORT_BREAK_EVEN is the size of the largest partition we should insertion sort.
604    Anything bigger and we use qsort. If you make this too small, the qsort
605    will probably break (or become less efficient), because it doesn't expect
606    the middle element of a partition to be the same as the right or left -
607    you have been warned).
608 */
609 #ifndef QSORT_BREAK_EVEN
610 #define QSORT_BREAK_EVEN 6
611 #endif
612 
613 /* QSORT_PLAY_SAFE is the size of the largest partition we're willing
614    to go quadratic on.  We innoculate larger partitions against
615    quadratic behavior by shuffling them before sorting.  This is not
616    an absolute guarantee of non-quadratic behavior, but it would take
617    staggeringly bad luck to pick extreme elements as the pivot
618    from randomized data.
619 */
620 #ifndef QSORT_PLAY_SAFE
621 #define QSORT_PLAY_SAFE 255
622 #endif
623 
624 /* ************************************************************* Data Types */
625 
626 /* hold left and right index values of a partition waiting to be sorted (the
627    partition includes both left and right - right is NOT one past the end or
628    anything like that).
629 */
630 struct partition_stack_entry {
631    int left;
632    int right;
633 #ifdef QSORT_ORDER_GUESS
634    int qsort_break_even;
635 #endif
636 };
637 
638 /* ******************************************************* Shorthand Macros */
639 
640 /* Note that these macros will be used from inside the qsort function where
641    we happen to know that the variable 'elt_size' contains the size of an
642    array element and the variable 'temp' points to enough space to hold a
643    temp element and the variable 'array' points to the array being sorted
644    and 'compare' is the pointer to the compare routine.
645 
646    Also note that there are very many highly architecture specific ways
647    these might be sped up, but this is simply the most generally portable
648    code I could think of.
649 */
650 
651 /* Return < 0 == 0 or > 0 as the value of elt1 is < elt2, == elt2, > elt2
652 */
653 #define qsort_cmp(elt1, elt2) \
654    ((*compare)(aTHX_ array[elt1], array[elt2]))
655 
656 #ifdef QSORT_ORDER_GUESS
657 #define QSORT_NOTICE_SWAP swapped++;
658 #else
659 #define QSORT_NOTICE_SWAP
660 #endif
661 
662 /* swaps contents of array elements elt1, elt2.
663 */
664 #define qsort_swap(elt1, elt2) \
665    STMT_START { \
666       QSORT_NOTICE_SWAP \
667       temp = array[elt1]; \
668       array[elt1] = array[elt2]; \
669       array[elt2] = temp; \
670    } STMT_END
671 
672 /* rotate contents of elt1, elt2, elt3 such that elt1 gets elt2, elt2 gets
673    elt3 and elt3 gets elt1.
674 */
675 #define qsort_rotate(elt1, elt2, elt3) \
676    STMT_START { \
677       QSORT_NOTICE_SWAP \
678       temp = array[elt1]; \
679       array[elt1] = array[elt2]; \
680       array[elt2] = array[elt3]; \
681       array[elt3] = temp; \
682    } STMT_END
683 
684 /* ************************************************************ Debug stuff */
685 
686 #ifdef QSORT_DEBUG
687 
688 static void
689 break_here()
690 {
691    return; /* good place to set a breakpoint */
692 }
693 
694 #define qsort_assert(t) (void)( (t) || (break_here(), 0) )
695 
696 static void
697 doqsort_all_asserts(
698    void * array,
699    size_t num_elts,
700    size_t elt_size,
701    int (*compare)(const void * elt1, const void * elt2),
702    int pc_left, int pc_right, int u_left, int u_right)
703 {
704    int i;
705 
706    qsort_assert(pc_left <= pc_right);
707    qsort_assert(u_right < pc_left);
708    qsort_assert(pc_right < u_left);
709    for (i = u_right + 1; i < pc_left; ++i) {
710       qsort_assert(qsort_cmp(i, pc_left) < 0);
711    }
712    for (i = pc_left; i < pc_right; ++i) {
713       qsort_assert(qsort_cmp(i, pc_right) == 0);
714    }
715    for (i = pc_right + 1; i < u_left; ++i) {
716       qsort_assert(qsort_cmp(pc_right, i) < 0);
717    }
718 }
719 
720 #define qsort_all_asserts(PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT) \
721    doqsort_all_asserts(array, num_elts, elt_size, compare, \
722                  PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT)
723 
724 #else
725 
726 #define qsort_assert(t) ((void)0)
727 
728 #define qsort_all_asserts(PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT) ((void)0)
729 
730 #endif
731 
732 /* ****************************************************************** qsort */
733 
734 STATIC void /* the standard unstable (u) quicksort (qsort) */
735 S_qsortsvu(pTHX_ SV ** array, size_t num_elts, SVCOMPARE_t compare)
736 {
737    register SV * temp;
738 
739    struct partition_stack_entry partition_stack[QSORT_MAX_STACK];
740    int next_stack_entry = 0;
741 
742    int part_left;
743    int part_right;
744 #ifdef QSORT_ORDER_GUESS
745    int qsort_break_even;
746    int swapped;
747 #endif
748 
749    /* Make sure we actually have work to do.
750    */
751    if (num_elts <= 1) {
752       return;
753    }
754 
755    /* Innoculate large partitions against quadratic behavior */
756    if (num_elts > QSORT_PLAY_SAFE) {
757       register size_t n, j;
758       register SV **q;
759       for (n = num_elts, q = array; n > 1; ) {
760          j = (size_t)(n-- * Drand01());
761          temp = q[j];
762          q[j] = q[n];
763          q[n] = temp;
764       }
765    }
766 
767    /* Setup the initial partition definition and fall into the sorting loop
768    */
769    part_left = 0;
770    part_right = (int)(num_elts - 1);
771 #ifdef QSORT_ORDER_GUESS
772    qsort_break_even = QSORT_BREAK_EVEN;
773 #else
774 #define qsort_break_even QSORT_BREAK_EVEN
775 #endif
776    for ( ; ; ) {
777       if ((part_right - part_left) >= qsort_break_even) {
778          /* OK, this is gonna get hairy, so lets try to document all the
779             concepts and abbreviations and variables and what they keep
780             track of:
781 
782             pc: pivot chunk - the set of array elements we accumulate in the
783                 middle of the partition, all equal in value to the original
784                 pivot element selected. The pc is defined by:
785 
786                 pc_left - the leftmost array index of the pc
787                 pc_right - the rightmost array index of the pc
788 
789                 we start with pc_left == pc_right and only one element
790                 in the pivot chunk (but it can grow during the scan).
791 
792             u:  uncompared elements - the set of elements in the partition
793                 we have not yet compared to the pivot value. There are two
794                 uncompared sets during the scan - one to the left of the pc
795                 and one to the right.
796 
797                 u_right - the rightmost index of the left side's uncompared set
798                 u_left - the leftmost index of the right side's uncompared set
799 
800                 The leftmost index of the left sides's uncompared set
801                 doesn't need its own variable because it is always defined
802                 by the leftmost edge of the whole partition (part_left). The
803                 same goes for the rightmost edge of the right partition
804                 (part_right).
805 
806                 We know there are no uncompared elements on the left once we
807                 get u_right < part_left and no uncompared elements on the
808                 right once u_left > part_right. When both these conditions
809                 are met, we have completed the scan of the partition.
810 
811                 Any elements which are between the pivot chunk and the
812                 uncompared elements should be less than the pivot value on
813                 the left side and greater than the pivot value on the right
814                 side (in fact, the goal of the whole algorithm is to arrange
815                 for that to be true and make the groups of less-than and
816                 greater-then elements into new partitions to sort again).
817 
818             As you marvel at the complexity of the code and wonder why it
819             has to be so confusing. Consider some of the things this level
820             of confusion brings:
821 
822             Once I do a compare, I squeeze every ounce of juice out of it. I
823             never do compare calls I don't have to do, and I certainly never
824             do redundant calls.
825 
826             I also never swap any elements unless I can prove there is a
827             good reason. Many sort algorithms will swap a known value with
828             an uncompared value just to get things in the right place (or
829             avoid complexity :-), but that uncompared value, once it gets
830             compared, may then have to be swapped again. A lot of the
831             complexity of this code is due to the fact that it never swaps
832             anything except compared values, and it only swaps them when the
833             compare shows they are out of position.
834          */
835          int pc_left, pc_right;
836          int u_right, u_left;
837 
838          int s;
839 
840          pc_left = ((part_left + part_right) / 2);
841          pc_right = pc_left;
842          u_right = pc_left - 1;
843          u_left = pc_right + 1;
844 
845          /* Qsort works best when the pivot value is also the median value
846             in the partition (unfortunately you can't find the median value
847             without first sorting :-), so to give the algorithm a helping
848             hand, we pick 3 elements and sort them and use the median value
849             of that tiny set as the pivot value.
850 
851             Some versions of qsort like to use the left middle and right as
852             the 3 elements to sort so they can insure the ends of the
853             partition will contain values which will stop the scan in the
854             compare loop, but when you have to call an arbitrarily complex
855             routine to do a compare, its really better to just keep track of
856             array index values to know when you hit the edge of the
857             partition and avoid the extra compare. An even better reason to
858             avoid using a compare call is the fact that you can drop off the
859             edge of the array if someone foolishly provides you with an
860             unstable compare function that doesn't always provide consistent
861             results.
862 
863             So, since it is simpler for us to compare the three adjacent
864             elements in the middle of the partition, those are the ones we
865             pick here (conveniently pointed at by u_right, pc_left, and
866             u_left). The values of the left, center, and right elements
867             are refered to as l c and r in the following comments.
868          */
869 
870 #ifdef QSORT_ORDER_GUESS
871          swapped = 0;
872 #endif
873          s = qsort_cmp(u_right, pc_left);
874          if (s < 0) {
875             /* l < c */
876             s = qsort_cmp(pc_left, u_left);
877             /* if l < c, c < r - already in order - nothing to do */
878             if (s == 0) {
879                /* l < c, c == r - already in order, pc grows */
880                ++pc_right;
881                qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
882             } else if (s > 0) {
883                /* l < c, c > r - need to know more */
884                s = qsort_cmp(u_right, u_left);
885                if (s < 0) {
886                   /* l < c, c > r, l < r - swap c & r to get ordered */
887                   qsort_swap(pc_left, u_left);
888                   qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
889                } else if (s == 0) {
890                   /* l < c, c > r, l == r - swap c&r, grow pc */
891                   qsort_swap(pc_left, u_left);
892                   --pc_left;
893                   qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
894                } else {
895                   /* l < c, c > r, l > r - make lcr into rlc to get ordered */
896                   qsort_rotate(pc_left, u_right, u_left);
897                   qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
898                }
899             }
900          } else if (s == 0) {
901             /* l == c */
902             s = qsort_cmp(pc_left, u_left);
903             if (s < 0) {
904                /* l == c, c < r - already in order, grow pc */
905                --pc_left;
906                qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
907             } else if (s == 0) {
908                /* l == c, c == r - already in order, grow pc both ways */
909                --pc_left;
910                ++pc_right;
911                qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
912             } else {
913                /* l == c, c > r - swap l & r, grow pc */
914                qsort_swap(u_right, u_left);
915                ++pc_right;
916                qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
917             }
918          } else {
919             /* l > c */
920             s = qsort_cmp(pc_left, u_left);
921             if (s < 0) {
922                /* l > c, c < r - need to know more */
923                s = qsort_cmp(u_right, u_left);
924                if (s < 0) {
925                   /* l > c, c < r, l < r - swap l & c to get ordered */
926                   qsort_swap(u_right, pc_left);
927                   qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
928                } else if (s == 0) {
929                   /* l > c, c < r, l == r - swap l & c, grow pc */
930                   qsort_swap(u_right, pc_left);
931                   ++pc_right;
932                   qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
933                } else {
934                   /* l > c, c < r, l > r - rotate lcr into crl to order */
935                   qsort_rotate(u_right, pc_left, u_left);
936                   qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
937                }
938             } else if (s == 0) {
939                /* l > c, c == r - swap ends, grow pc */
940                qsort_swap(u_right, u_left);
941                --pc_left;
942                qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
943             } else {
944                /* l > c, c > r - swap ends to get in order */
945                qsort_swap(u_right, u_left);
946                qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
947             }
948          }
949          /* We now know the 3 middle elements have been compared and
950             arranged in the desired order, so we can shrink the uncompared
951             sets on both sides
952          */
953          --u_right;
954          ++u_left;
955          qsort_all_asserts(pc_left, pc_right, u_left, u_right);
956 
957          /* The above massive nested if was the simple part :-). We now have
958             the middle 3 elements ordered and we need to scan through the
959             uncompared sets on either side, swapping elements that are on
960             the wrong side or simply shuffling equal elements around to get
961             all equal elements into the pivot chunk.
962          */
963 
964          for ( ; ; ) {
965             int still_work_on_left;
966             int still_work_on_right;
967 
968             /* Scan the uncompared values on the left. If I find a value
969                equal to the pivot value, move it over so it is adjacent to
970                the pivot chunk and expand the pivot chunk. If I find a value
971                less than the pivot value, then just leave it - its already
972                on the correct side of the partition. If I find a greater
973                value, then stop the scan.
974             */
975             while ((still_work_on_left = (u_right >= part_left))) {
976                s = qsort_cmp(u_right, pc_left);
977                if (s < 0) {
978                   --u_right;
979                } else if (s == 0) {
980                   --pc_left;
981                   if (pc_left != u_right) {
982                      qsort_swap(u_right, pc_left);
983                   }
984                   --u_right;
985                } else {
986                   break;
987                }
988                qsort_assert(u_right < pc_left);
989                qsort_assert(pc_left <= pc_right);
990                qsort_assert(qsort_cmp(u_right + 1, pc_left) <= 0);
991                qsort_assert(qsort_cmp(pc_left, pc_right) == 0);
992             }
993 
994             /* Do a mirror image scan of uncompared values on the right
995             */
996             while ((still_work_on_right = (u_left <= part_right))) {
997                s = qsort_cmp(pc_right, u_left);
998                if (s < 0) {
999                   ++u_left;
1000                } else if (s == 0) {
1001                   ++pc_right;
1002                   if (pc_right != u_left) {
1003                      qsort_swap(pc_right, u_left);
1004                   }
1005                   ++u_left;
1006                } else {
1007                   break;
1008                }
1009                qsort_assert(u_left > pc_right);
1010                qsort_assert(pc_left <= pc_right);
1011                qsort_assert(qsort_cmp(pc_right, u_left - 1) <= 0);
1012                qsort_assert(qsort_cmp(pc_left, pc_right) == 0);
1013             }
1014 
1015             if (still_work_on_left) {
1016                /* I know I have a value on the left side which needs to be
1017                   on the right side, but I need to know more to decide
1018                   exactly the best thing to do with it.
1019                */
1020                if (still_work_on_right) {
1021                   /* I know I have values on both side which are out of
1022                      position. This is a big win because I kill two birds
1023                      with one swap (so to speak). I can advance the
1024                      uncompared pointers on both sides after swapping both
1025                      of them into the right place.
1026                   */
1027                   qsort_swap(u_right, u_left);
1028                   --u_right;
1029                   ++u_left;
1030                   qsort_all_asserts(pc_left, pc_right, u_left, u_right);
1031                } else {
1032                   /* I have an out of position value on the left, but the
1033                      right is fully scanned, so I "slide" the pivot chunk
1034                      and any less-than values left one to make room for the
1035                      greater value over on the right. If the out of position
1036                      value is immediately adjacent to the pivot chunk (there
1037                      are no less-than values), I can do that with a swap,
1038                      otherwise, I have to rotate one of the less than values
1039                      into the former position of the out of position value
1040                      and the right end of the pivot chunk into the left end
1041                      (got all that?).
1042                   */
1043                   --pc_left;
1044                   if (pc_left == u_right) {
1045                      qsort_swap(u_right, pc_right);
1046                      qsort_all_asserts(pc_left, pc_right-1, u_left, u_right-1);
1047                   } else {
1048                      qsort_rotate(u_right, pc_left, pc_right);
1049                      qsort_all_asserts(pc_left, pc_right-1, u_left, u_right-1);
1050                   }
1051                   --pc_right;
1052                   --u_right;
1053                }
1054             } else if (still_work_on_right) {
1055                /* Mirror image of complex case above: I have an out of
1056                   position value on the right, but the left is fully
1057                   scanned, so I need to shuffle things around to make room
1058                   for the right value on the left.
1059                */
1060                ++pc_right;
1061                if (pc_right == u_left) {
1062                   qsort_swap(u_left, pc_left);
1063                   qsort_all_asserts(pc_left+1, pc_right, u_left+1, u_right);
1064                } else {
1065                   qsort_rotate(pc_right, pc_left, u_left);
1066                   qsort_all_asserts(pc_left+1, pc_right, u_left+1, u_right);
1067                }
1068                ++pc_left;
1069                ++u_left;
1070             } else {
1071                /* No more scanning required on either side of partition,
1072                   break out of loop and figure out next set of partitions
1073                */
1074                break;
1075             }
1076          }
1077 
1078          /* The elements in the pivot chunk are now in the right place. They
1079             will never move or be compared again. All I have to do is decide
1080             what to do with the stuff to the left and right of the pivot
1081             chunk.
1082 
1083             Notes on the QSORT_ORDER_GUESS ifdef code:
1084 
1085             1. If I just built these partitions without swapping any (or
1086                very many) elements, there is a chance that the elements are
1087                already ordered properly (being properly ordered will
1088                certainly result in no swapping, but the converse can't be
1089                proved :-).
1090 
1091             2. A (properly written) insertion sort will run faster on
1092                already ordered data than qsort will.
1093 
1094             3. Perhaps there is some way to make a good guess about
1095                switching to an insertion sort earlier than partition size 6
1096                (for instance - we could save the partition size on the stack
1097                and increase the size each time we find we didn't swap, thus
1098                switching to insertion sort earlier for partitions with a
1099                history of not swapping).
1100 
1101             4. Naturally, if I just switch right away, it will make
1102                artificial benchmarks with pure ascending (or descending)
1103                data look really good, but is that a good reason in general?
1104                Hard to say...
1105          */
1106 
1107 #ifdef QSORT_ORDER_GUESS
1108          if (swapped < 3) {
1109 #if QSORT_ORDER_GUESS == 1
1110             qsort_break_even = (part_right - part_left) + 1;
1111 #endif
1112 #if QSORT_ORDER_GUESS == 2
1113             qsort_break_even *= 2;
1114 #endif
1115 #if QSORT_ORDER_GUESS == 3
1116             int prev_break = qsort_break_even;
1117             qsort_break_even *= qsort_break_even;
1118             if (qsort_break_even < prev_break) {
1119                qsort_break_even = (part_right - part_left) + 1;
1120             }
1121 #endif
1122          } else {
1123             qsort_break_even = QSORT_BREAK_EVEN;
1124          }
1125 #endif
1126 
1127          if (part_left < pc_left) {
1128             /* There are elements on the left which need more processing.
1129                Check the right as well before deciding what to do.
1130             */
1131             if (pc_right < part_right) {
1132                /* We have two partitions to be sorted. Stack the biggest one
1133                   and process the smallest one on the next iteration. This
1134                   minimizes the stack height by insuring that any additional
1135                   stack entries must come from the smallest partition which
1136                   (because it is smallest) will have the fewest
1137                   opportunities to generate additional stack entries.
1138                */
1139                if ((part_right - pc_right) > (pc_left - part_left)) {
1140                   /* stack the right partition, process the left */
1141                   partition_stack[next_stack_entry].left = pc_right + 1;
1142                   partition_stack[next_stack_entry].right = part_right;
1143 #ifdef QSORT_ORDER_GUESS
1144                   partition_stack[next_stack_entry].qsort_break_even = qsort_break_even;
1145 #endif
1146                   part_right = pc_left - 1;
1147                } else {
1148                   /* stack the left partition, process the right */
1149                   partition_stack[next_stack_entry].left = part_left;
1150                   partition_stack[next_stack_entry].right = pc_left - 1;
1151 #ifdef QSORT_ORDER_GUESS
1152                   partition_stack[next_stack_entry].qsort_break_even = qsort_break_even;
1153 #endif
1154                   part_left = pc_right + 1;
1155                }
1156                qsort_assert(next_stack_entry < QSORT_MAX_STACK);
1157                ++next_stack_entry;
1158             } else {
1159                /* The elements on the left are the only remaining elements
1160                   that need sorting, arrange for them to be processed as the
1161                   next partition.
1162                */
1163                part_right = pc_left - 1;
1164             }
1165          } else if (pc_right < part_right) {
1166             /* There is only one chunk on the right to be sorted, make it
1167                the new partition and loop back around.
1168             */
1169             part_left = pc_right + 1;
1170          } else {
1171             /* This whole partition wound up in the pivot chunk, so
1172                we need to get a new partition off the stack.
1173             */
1174             if (next_stack_entry == 0) {
1175                /* the stack is empty - we are done */
1176                break;
1177             }
1178             --next_stack_entry;
1179             part_left = partition_stack[next_stack_entry].left;
1180             part_right = partition_stack[next_stack_entry].right;
1181 #ifdef QSORT_ORDER_GUESS
1182             qsort_break_even = partition_stack[next_stack_entry].qsort_break_even;
1183 #endif
1184          }
1185       } else {
1186          /* This partition is too small to fool with qsort complexity, just
1187             do an ordinary insertion sort to minimize overhead.
1188          */
1189          int i;
1190          /* Assume 1st element is in right place already, and start checking
1191             at 2nd element to see where it should be inserted.
1192          */
1193          for (i = part_left + 1; i <= part_right; ++i) {
1194             int j;
1195             /* Scan (backwards - just in case 'i' is already in right place)
1196                through the elements already sorted to see if the ith element
1197                belongs ahead of one of them.
1198             */
1199             for (j = i - 1; j >= part_left; --j) {
1200                if (qsort_cmp(i, j) >= 0) {
1201                   /* i belongs right after j
1202                   */
1203                   break;
1204                }
1205             }
1206             ++j;
1207             if (j != i) {
1208                /* Looks like we really need to move some things
1209                */
1210 	       int k;
1211 	       temp = array[i];
1212 	       for (k = i - 1; k >= j; --k)
1213 		  array[k + 1] = array[k];
1214                array[j] = temp;
1215             }
1216          }
1217 
1218          /* That partition is now sorted, grab the next one, or get out
1219             of the loop if there aren't any more.
1220          */
1221 
1222          if (next_stack_entry == 0) {
1223             /* the stack is empty - we are done */
1224             break;
1225          }
1226          --next_stack_entry;
1227          part_left = partition_stack[next_stack_entry].left;
1228          part_right = partition_stack[next_stack_entry].right;
1229 #ifdef QSORT_ORDER_GUESS
1230          qsort_break_even = partition_stack[next_stack_entry].qsort_break_even;
1231 #endif
1232       }
1233    }
1234 
1235    /* Believe it or not, the array is sorted at this point! */
1236 }
1237 
1238 /* Stabilize what is, presumably, an otherwise unstable sort method.
1239  * We do that by allocating (or having on hand) an array of pointers
1240  * that is the same size as the original array of elements to be sorted.
1241  * We initialize this parallel array with the addresses of the original
1242  * array elements.  This indirection can make you crazy.
1243  * Some pictures can help.  After initializing, we have
1244  *
1245  *  indir                  list1
1246  * +----+                 +----+
1247  * |    | --------------> |    | ------> first element to be sorted
1248  * +----+                 +----+
1249  * |    | --------------> |    | ------> second element to be sorted
1250  * +----+                 +----+
1251  * |    | --------------> |    | ------> third element to be sorted
1252  * +----+                 +----+
1253  *  ...
1254  * +----+                 +----+
1255  * |    | --------------> |    | ------> n-1st element to be sorted
1256  * +----+                 +----+
1257  * |    | --------------> |    | ------> n-th element to be sorted
1258  * +----+                 +----+
1259  *
1260  * During the sort phase, we leave the elements of list1 where they are,
1261  * and sort the pointers in the indirect array in the same order determined
1262  * by the original comparison routine on the elements pointed to.
1263  * Because we don't move the elements of list1 around through
1264  * this phase, we can break ties on elements that compare equal
1265  * using their address in the list1 array, ensuring stabilty.
1266  * This leaves us with something looking like
1267  *
1268  *  indir                  list1
1269  * +----+                 +----+
1270  * |    | --+       +---> |    | ------> first element to be sorted
1271  * +----+   |       |     +----+
1272  * |    | --|-------|---> |    | ------> second element to be sorted
1273  * +----+   |       |     +----+
1274  * |    | --|-------+ +-> |    | ------> third element to be sorted
1275  * +----+   |         |   +----+
1276  *  ...
1277  * +----+    | |   | |    +----+
1278  * |    | ---|-+   | +--> |    | ------> n-1st element to be sorted
1279  * +----+    |     |      +----+
1280  * |    | ---+     +----> |    | ------> n-th element to be sorted
1281  * +----+                 +----+
1282  *
1283  * where the i-th element of the indirect array points to the element
1284  * that should be i-th in the sorted array.  After the sort phase,
1285  * we have to put the elements of list1 into the places
1286  * dictated by the indirect array.
1287  */
1288 
1289 
1290 static I32
1291 cmpindir(pTHX_ gptr a, gptr b)
1292 {
1293     I32 sense;
1294     gptr *ap = (gptr *)a;
1295     gptr *bp = (gptr *)b;
1296 
1297     if ((sense = PL_sort_RealCmp(aTHX_ *ap, *bp)) == 0)
1298 	 sense = (ap > bp) ? 1 : ((ap < bp) ? -1 : 0);
1299     return sense;
1300 }
1301 
1302 STATIC void
1303 S_qsortsv(pTHX_ gptr *list1, size_t nmemb, SVCOMPARE_t cmp)
1304 {
1305     SV *hintsv;
1306 
1307     if (SORTHINTS(hintsv) & HINT_SORT_STABLE) {
1308 	 register gptr **pp, *q;
1309 	 register size_t n, j, i;
1310 	 gptr *small[SMALLSORT], **indir, tmp;
1311 	 SVCOMPARE_t savecmp;
1312 	 if (nmemb <= 1) return;     /* sorted trivially */
1313 
1314 	 /* Small arrays can use the stack, big ones must be allocated */
1315 	 if (nmemb <= SMALLSORT) indir = small;
1316 	 else { New(1799, indir, nmemb, gptr *); }
1317 
1318 	 /* Copy pointers to original array elements into indirect array */
1319 	 for (n = nmemb, pp = indir, q = list1; n--; ) *pp++ = q++;
1320 
1321 	 savecmp = PL_sort_RealCmp;	/* Save current comparison routine, if any */
1322 	 PL_sort_RealCmp = cmp;	/* Put comparison routine where cmpindir can find it */
1323 
1324 	 /* sort, with indirection */
1325 	 S_qsortsvu(aTHX_ (gptr *)indir, nmemb, cmpindir);
1326 
1327 	 pp = indir;
1328 	 q = list1;
1329 	 for (n = nmemb; n--; ) {
1330 	      /* Assert A: all elements of q with index > n are already
1331 	       * in place.  This is vacuosly true at the start, and we
1332 	       * put element n where it belongs below (if it wasn't
1333 	       * already where it belonged). Assert B: we only move
1334 	       * elements that aren't where they belong,
1335 	       * so, by A, we never tamper with elements above n.
1336 	       */
1337 	      j = pp[n] - q;		/* This sets j so that q[j] is
1338 					 * at pp[n].  *pp[j] belongs in
1339 					 * q[j], by construction.
1340 					 */
1341 	      if (n != j) {		/* all's well if n == j */
1342 		   tmp = q[j];		/* save what's in q[j] */
1343 		   do {
1344 			q[j] = *pp[j];	/* put *pp[j] where it belongs */
1345 			i = pp[j] - q;	/* the index in q of the element
1346 					 * just moved */
1347 			pp[j] = q + j;	/* this is ok now */
1348 		   } while ((j = i) != n);
1349 		   /* There are only finitely many (nmemb) addresses
1350 		    * in the pp array.
1351 		    * So we must eventually revisit an index we saw before.
1352 		    * Suppose the first revisited index is k != n.
1353 		    * An index is visited because something else belongs there.
1354 		    * If we visit k twice, then two different elements must
1355 		    * belong in the same place, which cannot be.
1356 		    * So j must get back to n, the loop terminates,
1357 		    * and we put the saved element where it belongs.
1358 		    */
1359 		   q[n] = tmp;		/* put what belongs into
1360 					 * the n-th element */
1361 	      }
1362 	 }
1363 
1364 	/* free iff allocated */
1365 	 if (indir != small) { Safefree(indir); }
1366 	 /* restore prevailing comparison routine */
1367 	 PL_sort_RealCmp = savecmp;
1368     } else {
1369 	 S_qsortsvu(aTHX_ list1, nmemb, cmp);
1370     }
1371 }
1372 
1373 /*
1374 =head1 Array Manipulation Functions
1375 
1376 =for apidoc sortsv
1377 
1378 Sort an array. Here is an example:
1379 
1380     sortsv(AvARRAY(av), av_len(av)+1, Perl_sv_cmp_locale);
1381 
1382 See lib/sort.pm for details about controlling the sorting algorithm.
1383 
1384 =cut
1385 */
1386 
1387 void
1388 Perl_sortsv(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp)
1389 {
1390     void (*sortsvp)(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp) =
1391         S_mergesortsv;
1392     SV *hintsv;
1393     I32 hints;
1394 
1395     /*  Sun's Compiler (cc: WorkShop Compilers 4.2 30 Oct 1996 C 4.2) used
1396 	to miscompile this function under optimization -O.  If you get test
1397 	errors related to picking the correct sort() function, try recompiling
1398 	this file without optimiziation.  -- A.D.  4/2002.
1399     */
1400     hints = SORTHINTS(hintsv);
1401     if (hints & HINT_SORT_QUICKSORT) {
1402 	sortsvp = S_qsortsv;
1403     }
1404     else {
1405 	/* The default as of 5.8.0 is mergesort */
1406 	sortsvp = S_mergesortsv;
1407     }
1408 
1409     sortsvp(aTHX_ array, nmemb, cmp);
1410 }
1411 
1412 PP(pp_sort)
1413 {
1414     dSP; dMARK; dORIGMARK;
1415     register SV **up;
1416     SV **myorigmark = ORIGMARK;
1417     register I32 max;
1418     HV *stash;
1419     GV *gv;
1420     CV *cv = 0;
1421     I32 gimme = GIMME;
1422     OP* nextop = PL_op->op_next;
1423     I32 overloading = 0;
1424     bool hasargs = FALSE;
1425     I32 is_xsub = 0;
1426 
1427     if (gimme != G_ARRAY) {
1428 	SP = MARK;
1429 	RETPUSHUNDEF;
1430     }
1431 
1432     ENTER;
1433     SAVEVPTR(PL_sortcop);
1434     if (PL_op->op_flags & OPf_STACKED) {
1435 	if (PL_op->op_flags & OPf_SPECIAL) {
1436 	    OP *kid = cLISTOP->op_first->op_sibling;	/* pass pushmark */
1437 	    kid = kUNOP->op_first;			/* pass rv2gv */
1438 	    kid = kUNOP->op_first;			/* pass leave */
1439 	    PL_sortcop = kid->op_next;
1440 	    stash = CopSTASH(PL_curcop);
1441 	}
1442 	else {
1443 	    cv = sv_2cv(*++MARK, &stash, &gv, 0);
1444 	    if (cv && SvPOK(cv)) {
1445 		STRLEN n_a;
1446 		char *proto = SvPV((SV*)cv, n_a);
1447 		if (proto && strEQ(proto, "$$")) {
1448 		    hasargs = TRUE;
1449 		}
1450 	    }
1451 	    if (!(cv && CvROOT(cv))) {
1452 		if (cv && CvXSUB(cv)) {
1453 		    is_xsub = 1;
1454 		}
1455 		else if (gv) {
1456 		    SV *tmpstr = sv_newmortal();
1457 		    gv_efullname3(tmpstr, gv, Nullch);
1458 		    DIE(aTHX_ "Undefined sort subroutine \"%s\" called",
1459 			SvPVX(tmpstr));
1460 		}
1461 		else {
1462 		    DIE(aTHX_ "Undefined subroutine in sort");
1463 		}
1464 	    }
1465 
1466 	    if (is_xsub)
1467 		PL_sortcop = (OP*)cv;
1468 	    else {
1469 		PL_sortcop = CvSTART(cv);
1470 		SAVEVPTR(CvROOT(cv)->op_ppaddr);
1471 		CvROOT(cv)->op_ppaddr = PL_ppaddr[OP_NULL];
1472 
1473 		SAVEVPTR(PL_curpad);
1474 		PL_curpad = AvARRAY((AV*)AvARRAY(CvPADLIST(cv))[1]);
1475             }
1476 	}
1477     }
1478     else {
1479 	PL_sortcop = Nullop;
1480 	stash = CopSTASH(PL_curcop);
1481     }
1482 
1483     up = myorigmark + 1;
1484     while (MARK < SP) {	/* This may or may not shift down one here. */
1485 	/*SUPPRESS 560*/
1486 	if ((*up = *++MARK)) {			/* Weed out nulls. */
1487 	    SvTEMP_off(*up);
1488 	    if (!PL_sortcop && !SvPOK(*up)) {
1489 		STRLEN n_a;
1490 	        if (SvAMAGIC(*up))
1491 	            overloading = 1;
1492 	        else
1493 		    (void)sv_2pv(*up, &n_a);
1494 	    }
1495 	    up++;
1496 	}
1497     }
1498     max = --up - myorigmark;
1499     if (PL_sortcop) {
1500 	if (max > 1) {
1501 	    PERL_CONTEXT *cx;
1502 	    SV** newsp;
1503 	    bool oldcatch = CATCH_GET;
1504 
1505 	    SAVETMPS;
1506 	    SAVEOP();
1507 
1508 	    CATCH_SET(TRUE);
1509 	    PUSHSTACKi(PERLSI_SORT);
1510 	    if (!hasargs && !is_xsub) {
1511 		if (PL_sortstash != stash || !PL_firstgv || !PL_secondgv) {
1512 		    SAVESPTR(PL_firstgv);
1513 		    SAVESPTR(PL_secondgv);
1514 		    PL_firstgv = gv_fetchpv("a", TRUE, SVt_PV);
1515 		    PL_secondgv = gv_fetchpv("b", TRUE, SVt_PV);
1516 		    PL_sortstash = stash;
1517 		}
1518 #ifdef USE_5005THREADS
1519 		sv_lock((SV *)PL_firstgv);
1520 		sv_lock((SV *)PL_secondgv);
1521 #endif
1522 		SAVESPTR(GvSV(PL_firstgv));
1523 		SAVESPTR(GvSV(PL_secondgv));
1524 	    }
1525 
1526 	    PUSHBLOCK(cx, CXt_NULL, PL_stack_base);
1527 	    if (!(PL_op->op_flags & OPf_SPECIAL)) {
1528 		cx->cx_type = CXt_SUB;
1529 		cx->blk_gimme = G_SCALAR;
1530 		PUSHSUB(cx);
1531 		if (!CvDEPTH(cv))
1532 		    (void)SvREFCNT_inc(cv); /* in preparation for POPSUB */
1533 	    }
1534 	    PL_sortcxix = cxstack_ix;
1535 
1536 	    if (hasargs && !is_xsub) {
1537 		/* This is mostly copied from pp_entersub */
1538 		AV *av = (AV*)PL_curpad[0];
1539 
1540 #ifndef USE_5005THREADS
1541 		cx->blk_sub.savearray = GvAV(PL_defgv);
1542 		GvAV(PL_defgv) = (AV*)SvREFCNT_inc(av);
1543 #endif /* USE_5005THREADS */
1544 		cx->blk_sub.oldcurpad = PL_curpad;
1545 		cx->blk_sub.argarray = av;
1546 	    }
1547            sortsv((myorigmark+1), max,
1548                   is_xsub ? sortcv_xsub : hasargs ? sortcv_stacked : sortcv);
1549 
1550 	    POPBLOCK(cx,PL_curpm);
1551 	    PL_stack_sp = newsp;
1552 	    POPSTACK;
1553 	    CATCH_SET(oldcatch);
1554 	}
1555     }
1556     else {
1557 	if (max > 1) {
1558 	    MEXTEND(SP, 20);	/* Can't afford stack realloc on signal. */
1559 	    sortsv(ORIGMARK+1, max,
1560                   (PL_op->op_private & OPpSORT_NUMERIC)
1561 			? ( (PL_op->op_private & OPpSORT_INTEGER)
1562 			    ? ( overloading ? amagic_i_ncmp : sv_i_ncmp)
1563 			    : ( overloading ? amagic_ncmp : sv_ncmp))
1564 			: ( IN_LOCALE_RUNTIME
1565 			    ? ( overloading
1566 				? amagic_cmp_locale
1567 				: sv_cmp_locale_static)
1568 			    : ( overloading ? amagic_cmp : sv_cmp_static)));
1569 	    if (PL_op->op_private & OPpSORT_REVERSE) {
1570 		SV **p = ORIGMARK+1;
1571 		SV **q = ORIGMARK+max;
1572 		while (p < q) {
1573 		    SV *tmp = *p;
1574 		    *p++ = *q;
1575 		    *q-- = tmp;
1576 		}
1577 	    }
1578 	}
1579     }
1580     LEAVE;
1581     PL_stack_sp = ORIGMARK + max;
1582     return nextop;
1583 }
1584 
1585 static I32
1586 sortcv(pTHX_ SV *a, SV *b)
1587 {
1588     I32 oldsaveix = PL_savestack_ix;
1589     I32 oldscopeix = PL_scopestack_ix;
1590     I32 result;
1591     GvSV(PL_firstgv) = a;
1592     GvSV(PL_secondgv) = b;
1593     PL_stack_sp = PL_stack_base;
1594     PL_op = PL_sortcop;
1595     CALLRUNOPS(aTHX);
1596     if (PL_stack_sp != PL_stack_base + 1)
1597 	Perl_croak(aTHX_ "Sort subroutine didn't return single value");
1598     if (!SvNIOKp(*PL_stack_sp))
1599 	Perl_croak(aTHX_ "Sort subroutine didn't return a numeric value");
1600     result = SvIV(*PL_stack_sp);
1601     while (PL_scopestack_ix > oldscopeix) {
1602 	LEAVE;
1603     }
1604     leave_scope(oldsaveix);
1605     return result;
1606 }
1607 
1608 static I32
1609 sortcv_stacked(pTHX_ SV *a, SV *b)
1610 {
1611     I32 oldsaveix = PL_savestack_ix;
1612     I32 oldscopeix = PL_scopestack_ix;
1613     I32 result;
1614     AV *av;
1615 
1616 #ifdef USE_5005THREADS
1617     av = (AV*)PL_curpad[0];
1618 #else
1619     av = GvAV(PL_defgv);
1620 #endif
1621 
1622     if (AvMAX(av) < 1) {
1623 	SV** ary = AvALLOC(av);
1624 	if (AvARRAY(av) != ary) {
1625 	    AvMAX(av) += AvARRAY(av) - AvALLOC(av);
1626 	    SvPVX(av) = (char*)ary;
1627 	}
1628 	if (AvMAX(av) < 1) {
1629 	    AvMAX(av) = 1;
1630 	    Renew(ary,2,SV*);
1631 	    SvPVX(av) = (char*)ary;
1632 	}
1633     }
1634     AvFILLp(av) = 1;
1635 
1636     AvARRAY(av)[0] = a;
1637     AvARRAY(av)[1] = b;
1638     PL_stack_sp = PL_stack_base;
1639     PL_op = PL_sortcop;
1640     CALLRUNOPS(aTHX);
1641     if (PL_stack_sp != PL_stack_base + 1)
1642 	Perl_croak(aTHX_ "Sort subroutine didn't return single value");
1643     if (!SvNIOKp(*PL_stack_sp))
1644 	Perl_croak(aTHX_ "Sort subroutine didn't return a numeric value");
1645     result = SvIV(*PL_stack_sp);
1646     while (PL_scopestack_ix > oldscopeix) {
1647 	LEAVE;
1648     }
1649     leave_scope(oldsaveix);
1650     return result;
1651 }
1652 
1653 static I32
1654 sortcv_xsub(pTHX_ SV *a, SV *b)
1655 {
1656     dSP;
1657     I32 oldsaveix = PL_savestack_ix;
1658     I32 oldscopeix = PL_scopestack_ix;
1659     I32 result;
1660     CV *cv=(CV*)PL_sortcop;
1661 
1662     SP = PL_stack_base;
1663     PUSHMARK(SP);
1664     EXTEND(SP, 2);
1665     *++SP = a;
1666     *++SP = b;
1667     PUTBACK;
1668     (void)(*CvXSUB(cv))(aTHX_ cv);
1669     if (PL_stack_sp != PL_stack_base + 1)
1670 	Perl_croak(aTHX_ "Sort subroutine didn't return single value");
1671     if (!SvNIOKp(*PL_stack_sp))
1672 	Perl_croak(aTHX_ "Sort subroutine didn't return a numeric value");
1673     result = SvIV(*PL_stack_sp);
1674     while (PL_scopestack_ix > oldscopeix) {
1675 	LEAVE;
1676     }
1677     leave_scope(oldsaveix);
1678     return result;
1679 }
1680 
1681 
1682 static I32
1683 sv_ncmp(pTHX_ SV *a, SV *b)
1684 {
1685     NV nv1 = SvNV(a);
1686     NV nv2 = SvNV(b);
1687     return nv1 < nv2 ? -1 : nv1 > nv2 ? 1 : 0;
1688 }
1689 
1690 static I32
1691 sv_i_ncmp(pTHX_ SV *a, SV *b)
1692 {
1693     IV iv1 = SvIV(a);
1694     IV iv2 = SvIV(b);
1695     return iv1 < iv2 ? -1 : iv1 > iv2 ? 1 : 0;
1696 }
1697 #define tryCALL_AMAGICbin(left,right,meth,svp) STMT_START { \
1698 	  *svp = Nullsv;				\
1699           if (PL_amagic_generation) { \
1700 	    if (SvAMAGIC(left)||SvAMAGIC(right))\
1701 		*svp = amagic_call(left, \
1702 				   right, \
1703 				   CAT2(meth,_amg), \
1704 				   0); \
1705 	  } \
1706 	} STMT_END
1707 
1708 static I32
1709 amagic_ncmp(pTHX_ register SV *a, register SV *b)
1710 {
1711     SV *tmpsv;
1712     tryCALL_AMAGICbin(a,b,ncmp,&tmpsv);
1713     if (tmpsv) {
1714     	NV d;
1715 
1716         if (SvIOK(tmpsv)) {
1717             I32 i = SvIVX(tmpsv);
1718             if (i > 0)
1719                return 1;
1720             return i? -1 : 0;
1721         }
1722         d = SvNV(tmpsv);
1723         if (d > 0)
1724            return 1;
1725         return d? -1 : 0;
1726      }
1727      return sv_ncmp(aTHX_ a, b);
1728 }
1729 
1730 static I32
1731 amagic_i_ncmp(pTHX_ register SV *a, register SV *b)
1732 {
1733     SV *tmpsv;
1734     tryCALL_AMAGICbin(a,b,ncmp,&tmpsv);
1735     if (tmpsv) {
1736     	NV d;
1737 
1738         if (SvIOK(tmpsv)) {
1739             I32 i = SvIVX(tmpsv);
1740             if (i > 0)
1741                return 1;
1742             return i? -1 : 0;
1743         }
1744         d = SvNV(tmpsv);
1745         if (d > 0)
1746            return 1;
1747         return d? -1 : 0;
1748     }
1749     return sv_i_ncmp(aTHX_ a, b);
1750 }
1751 
1752 static I32
1753 amagic_cmp(pTHX_ register SV *str1, register SV *str2)
1754 {
1755     SV *tmpsv;
1756     tryCALL_AMAGICbin(str1,str2,scmp,&tmpsv);
1757     if (tmpsv) {
1758     	NV d;
1759 
1760         if (SvIOK(tmpsv)) {
1761             I32 i = SvIVX(tmpsv);
1762             if (i > 0)
1763                return 1;
1764             return i? -1 : 0;
1765         }
1766         d = SvNV(tmpsv);
1767         if (d > 0)
1768            return 1;
1769         return d? -1 : 0;
1770     }
1771     return sv_cmp(str1, str2);
1772 }
1773 
1774 static I32
1775 amagic_cmp_locale(pTHX_ register SV *str1, register SV *str2)
1776 {
1777     SV *tmpsv;
1778     tryCALL_AMAGICbin(str1,str2,scmp,&tmpsv);
1779     if (tmpsv) {
1780     	NV d;
1781 
1782         if (SvIOK(tmpsv)) {
1783             I32 i = SvIVX(tmpsv);
1784             if (i > 0)
1785                return 1;
1786             return i? -1 : 0;
1787         }
1788         d = SvNV(tmpsv);
1789         if (d > 0)
1790            return 1;
1791         return d? -1 : 0;
1792     }
1793     return sv_cmp_locale(str1, str2);
1794 }
1795