xref: /openbsd-src/gnu/usr.bin/perl/pp_sort.c (revision 8500990981f885cbe5e6a4958549cacc238b5ae6)
1 /*    pp_sort.c
2  *
3  *    Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
4  *    2000, 2001, 2002, 2003, by Larry Wall and others
5  *
6  *    You may distribute under the terms of either the GNU General Public
7  *    License or the Artistic License, as specified in the README file.
8  *
9  */
10 
11 /*
12  *   ...they shuffled back towards the rear of the line. 'No, not at the
13  *   rear!'  the slave-driver shouted. 'Three files up. And stay there...
14  */
15 
16 #include "EXTERN.h"
17 #define PERL_IN_PP_SORT_C
18 #include "perl.h"
19 
20 #if defined(UNDER_CE)
21 /* looks like 'small' is reserved word for WINCE (or somesuch)*/
22 #define	small xsmall
23 #endif
24 
25 static I32 sortcv(pTHX_ SV *a, SV *b);
26 static I32 sortcv_stacked(pTHX_ SV *a, SV *b);
27 static I32 sortcv_xsub(pTHX_ SV *a, SV *b);
28 static I32 sv_ncmp(pTHX_ SV *a, SV *b);
29 static I32 sv_i_ncmp(pTHX_ SV *a, SV *b);
30 static I32 amagic_ncmp(pTHX_ SV *a, SV *b);
31 static I32 amagic_i_ncmp(pTHX_ SV *a, SV *b);
32 static I32 amagic_cmp(pTHX_ SV *a, SV *b);
33 static I32 amagic_cmp_locale(pTHX_ SV *a, SV *b);
34 
35 #define sv_cmp_static Perl_sv_cmp
36 #define sv_cmp_locale_static Perl_sv_cmp_locale
37 
38 #define SORTHINTS(hintsv) \
39     (((hintsv) = GvSV(gv_fetchpv("sort::hints", GV_ADDMULTI, SVt_IV))), \
40     (SvIOK(hintsv) ? ((I32)SvIV(hintsv)) : 0))
41 
42 #ifndef SMALLSORT
43 #define	SMALLSORT (200)
44 #endif
45 
46 /*
47  * The mergesort implementation is by Peter M. Mcilroy <pmcilroy@lucent.com>.
48  *
49  * The original code was written in conjunction with BSD Computer Software
50  * Research Group at University of California, Berkeley.
51  *
52  * See also: "Optimistic Merge Sort" (SODA '92)
53  *
54  * The integration to Perl is by John P. Linderman <jpl@research.att.com>.
55  *
56  * The code can be distributed under the same terms as Perl itself.
57  *
58  */
59 
60 
61 typedef char * aptr;		/* pointer for arithmetic on sizes */
62 typedef SV * gptr;		/* pointers in our lists */
63 
64 /* Binary merge internal sort, with a few special mods
65 ** for the special perl environment it now finds itself in.
66 **
67 ** Things that were once options have been hotwired
68 ** to values suitable for this use.  In particular, we'll always
69 ** initialize looking for natural runs, we'll always produce stable
70 ** output, and we'll always do Peter McIlroy's binary merge.
71 */
72 
73 /* Pointer types for arithmetic and storage and convenience casts */
74 
75 #define	APTR(P)	((aptr)(P))
76 #define	GPTP(P)	((gptr *)(P))
77 #define GPPP(P) ((gptr **)(P))
78 
79 
80 /* byte offset from pointer P to (larger) pointer Q */
81 #define	BYTEOFF(P, Q) (APTR(Q) - APTR(P))
82 
83 #define PSIZE sizeof(gptr)
84 
85 /* If PSIZE is power of 2, make PSHIFT that power, if that helps */
86 
87 #ifdef	PSHIFT
88 #define	PNELEM(P, Q)	(BYTEOFF(P,Q) >> (PSHIFT))
89 #define	PNBYTE(N)	((N) << (PSHIFT))
90 #define	PINDEX(P, N)	(GPTP(APTR(P) + PNBYTE(N)))
91 #else
92 /* Leave optimization to compiler */
93 #define	PNELEM(P, Q)	(GPTP(Q) - GPTP(P))
94 #define	PNBYTE(N)	((N) * (PSIZE))
95 #define	PINDEX(P, N)	(GPTP(P) + (N))
96 #endif
97 
98 /* Pointer into other corresponding to pointer into this */
99 #define	POTHER(P, THIS, OTHER) GPTP(APTR(OTHER) + BYTEOFF(THIS,P))
100 
101 #define FROMTOUPTO(src, dst, lim) do *dst++ = *src++; while(src<lim)
102 
103 
104 /* Runs are identified by a pointer in the auxilliary list.
105 ** The pointer is at the start of the list,
106 ** and it points to the start of the next list.
107 ** NEXT is used as an lvalue, too.
108 */
109 
110 #define	NEXT(P)		(*GPPP(P))
111 
112 
113 /* PTHRESH is the minimum number of pairs with the same sense to justify
114 ** checking for a run and extending it.  Note that PTHRESH counts PAIRS,
115 ** not just elements, so PTHRESH == 8 means a run of 16.
116 */
117 
118 #define	PTHRESH (8)
119 
120 /* RTHRESH is the number of elements in a run that must compare low
121 ** to the low element from the opposing run before we justify
122 ** doing a binary rampup instead of single stepping.
123 ** In random input, N in a row low should only happen with
124 ** probability 2^(1-N), so we can risk that we are dealing
125 ** with orderly input without paying much when we aren't.
126 */
127 
128 #define RTHRESH (6)
129 
130 
131 /*
132 ** Overview of algorithm and variables.
133 ** The array of elements at list1 will be organized into runs of length 2,
134 ** or runs of length >= 2 * PTHRESH.  We only try to form long runs when
135 ** PTHRESH adjacent pairs compare in the same way, suggesting overall order.
136 **
137 ** Unless otherwise specified, pair pointers address the first of two elements.
138 **
139 ** b and b+1 are a pair that compare with sense ``sense''.
140 ** b is the ``bottom'' of adjacent pairs that might form a longer run.
141 **
142 ** p2 parallels b in the list2 array, where runs are defined by
143 ** a pointer chain.
144 **
145 ** t represents the ``top'' of the adjacent pairs that might extend
146 ** the run beginning at b.  Usually, t addresses a pair
147 ** that compares with opposite sense from (b,b+1).
148 ** However, it may also address a singleton element at the end of list1,
149 ** or it may be equal to ``last'', the first element beyond list1.
150 **
151 ** r addresses the Nth pair following b.  If this would be beyond t,
152 ** we back it off to t.  Only when r is less than t do we consider the
153 ** run long enough to consider checking.
154 **
155 ** q addresses a pair such that the pairs at b through q already form a run.
156 ** Often, q will equal b, indicating we only are sure of the pair itself.
157 ** However, a search on the previous cycle may have revealed a longer run,
158 ** so q may be greater than b.
159 **
160 ** p is used to work back from a candidate r, trying to reach q,
161 ** which would mean b through r would be a run.  If we discover such a run,
162 ** we start q at r and try to push it further towards t.
163 ** If b through r is NOT a run, we detect the wrong order at (p-1,p).
164 ** In any event, after the check (if any), we have two main cases.
165 **
166 ** 1) Short run.  b <= q < p <= r <= t.
167 **	b through q is a run (perhaps trivial)
168 **	q through p are uninteresting pairs
169 **	p through r is a run
170 **
171 ** 2) Long run.  b < r <= q < t.
172 **	b through q is a run (of length >= 2 * PTHRESH)
173 **
174 ** Note that degenerate cases are not only possible, but likely.
175 ** For example, if the pair following b compares with opposite sense,
176 ** then b == q < p == r == t.
177 */
178 
179 
180 static IV
181 dynprep(pTHX_ gptr *list1, gptr *list2, size_t nmemb, SVCOMPARE_t cmp)
182 {
183     I32 sense;
184     register gptr *b, *p, *q, *t, *p2;
185     register gptr c, *last, *r;
186     gptr *savep;
187     IV runs = 0;
188 
189     b = list1;
190     last = PINDEX(b, nmemb);
191     sense = (cmp(aTHX_ *b, *(b+1)) > 0);
192     for (p2 = list2; b < last; ) {
193 	/* We just started, or just reversed sense.
194 	** Set t at end of pairs with the prevailing sense.
195 	*/
196 	for (p = b+2, t = p; ++p < last; t = ++p) {
197 	    if ((cmp(aTHX_ *t, *p) > 0) != sense) break;
198 	}
199 	q = b;
200 	/* Having laid out the playing field, look for long runs */
201 	do {
202 	    p = r = b + (2 * PTHRESH);
203 	    if (r >= t) p = r = t;	/* too short to care about */
204 	    else {
205 		while (((cmp(aTHX_ *(p-1), *p) > 0) == sense) &&
206 		       ((p -= 2) > q));
207 		if (p <= q) {
208 		    /* b through r is a (long) run.
209 		    ** Extend it as far as possible.
210 		    */
211 		    p = q = r;
212 		    while (((p += 2) < t) &&
213 			   ((cmp(aTHX_ *(p-1), *p) > 0) == sense)) q = p;
214 		    r = p = q + 2;	/* no simple pairs, no after-run */
215 		}
216 	    }
217 	    if (q > b) {		/* run of greater than 2 at b */
218 		savep = p;
219 		p = q += 2;
220 		/* pick up singleton, if possible */
221 		if ((p == t) &&
222 		    ((t + 1) == last) &&
223 		    ((cmp(aTHX_ *(p-1), *p) > 0) == sense))
224 		    savep = r = p = q = last;
225 		p2 = NEXT(p2) = p2 + (p - b); ++runs;
226 		if (sense) while (b < --p) {
227 		    c = *b;
228 		    *b++ = *p;
229 		    *p = c;
230 		}
231 		p = savep;
232 	    }
233 	    while (q < p) {		/* simple pairs */
234 		p2 = NEXT(p2) = p2 + 2; ++runs;
235 		if (sense) {
236 		    c = *q++;
237 		    *(q-1) = *q;
238 		    *q++ = c;
239 		} else q += 2;
240 	    }
241 	    if (((b = p) == t) && ((t+1) == last)) {
242 		NEXT(p2) = p2 + 1; ++runs;
243 		b++;
244 	    }
245 	    q = r;
246 	} while (b < t);
247 	sense = !sense;
248     }
249     return runs;
250 }
251 
252 
253 /* The original merge sort, in use since 5.7, was as fast as, or faster than,
254  * qsort on many platforms, but slower than qsort, conspicuously so,
255  * on others.  The most likely explanation was platform-specific
256  * differences in cache sizes and relative speeds.
257  *
258  * The quicksort divide-and-conquer algorithm guarantees that, as the
259  * problem is subdivided into smaller and smaller parts, the parts
260  * fit into smaller (and faster) caches.  So it doesn't matter how
261  * many levels of cache exist, quicksort will "find" them, and,
262  * as long as smaller is faster, take advanatge of them.
263  *
264  * By contrast, consider how the original mergesort algorithm worked.
265  * Suppose we have five runs (each typically of length 2 after dynprep).
266  *
267  * pass               base                        aux
268  *  0              1 2 3 4 5
269  *  1                                           12 34 5
270  *  2                1234 5
271  *  3                                            12345
272  *  4                 12345
273  *
274  * Adjacent pairs are merged in "grand sweeps" through the input.
275  * This means, on pass 1, the records in runs 1 and 2 aren't revisited until
276  * runs 3 and 4 are merged and the runs from run 5 have been copied.
277  * The only cache that matters is one large enough to hold *all* the input.
278  * On some platforms, this may be many times slower than smaller caches.
279  *
280  * The following pseudo-code uses the same basic merge algorithm,
281  * but in a divide-and-conquer way.
282  *
283  * # merge $runs runs at offset $offset of list $list1 into $list2.
284  * # all unmerged runs ($runs == 1) originate in list $base.
285  * sub mgsort2 {
286  *     my ($offset, $runs, $base, $list1, $list2) = @_;
287  *
288  *     if ($runs == 1) {
289  *         if ($list1 is $base) copy run to $list2
290  *         return offset of end of list (or copy)
291  *     } else {
292  *         $off2 = mgsort2($offset, $runs-($runs/2), $base, $list2, $list1)
293  *         mgsort2($off2, $runs/2, $base, $list2, $list1)
294  *         merge the adjacent runs at $offset of $list1 into $list2
295  *         return the offset of the end of the merged runs
296  *     }
297  * }
298  * mgsort2(0, $runs, $base, $aux, $base);
299  *
300  * For our 5 runs, the tree of calls looks like
301  *
302  *           5
303  *      3        2
304  *   2     1   1   1
305  * 1   1
306  *
307  * 1   2   3   4   5
308  *
309  * and the corresponding activity looks like
310  *
311  * copy runs 1 and 2 from base to aux
312  * merge runs 1 and 2 from aux to base
313  * (run 3 is where it belongs, no copy needed)
314  * merge runs 12 and 3 from base to aux
315  * (runs 4 and 5 are where they belong, no copy needed)
316  * merge runs 4 and 5 from base to aux
317  * merge runs 123 and 45 from aux to base
318  *
319  * Note that we merge runs 1 and 2 immediately after copying them,
320  * while they are still likely to be in fast cache.  Similarly,
321  * run 3 is merged with run 12 while it still may be lingering in cache.
322  * This implementation should therefore enjoy much of the cache-friendly
323  * behavior that quicksort does.  In addition, it does less copying
324  * than the original mergesort implementation (only runs 1 and 2 are copied)
325  * and the "balancing" of merges is better (merged runs comprise more nearly
326  * equal numbers of original runs).
327  *
328  * The actual cache-friendly implementation will use a pseudo-stack
329  * to avoid recursion, and will unroll processing of runs of length 2,
330  * but it is otherwise similar to the recursive implementation.
331  */
332 
333 typedef struct {
334     IV	offset;		/* offset of 1st of 2 runs at this level */
335     IV	runs;		/* how many runs must be combined into 1 */
336 } off_runs;		/* pseudo-stack element */
337 
338 STATIC void
339 S_mergesortsv(pTHX_ gptr *base, size_t nmemb, SVCOMPARE_t cmp)
340 {
341     IV i, run, runs, offset;
342     I32 sense, level;
343     int iwhich;
344     register gptr *f1, *f2, *t, *b, *p, *tp2, *l1, *l2, *q;
345     gptr *aux, *list1, *list2;
346     gptr *p1;
347     gptr small[SMALLSORT];
348     gptr *which[3];
349     off_runs stack[60], *stackp;
350 
351     if (nmemb <= 1) return;			/* sorted trivially */
352     if (nmemb <= SMALLSORT) aux = small;	/* use stack for aux array */
353     else { New(799,aux,nmemb,gptr); }		/* allocate auxilliary array */
354     level = 0;
355     stackp = stack;
356     stackp->runs = dynprep(aTHX_ base, aux, nmemb, cmp);
357     stackp->offset = offset = 0;
358     which[0] = which[2] = base;
359     which[1] = aux;
360     for (;;) {
361 	/* On levels where both runs have be constructed (stackp->runs == 0),
362 	 * merge them, and note the offset of their end, in case the offset
363 	 * is needed at the next level up.  Hop up a level, and,
364 	 * as long as stackp->runs is 0, keep merging.
365 	 */
366 	if ((runs = stackp->runs) == 0) {
367 	    iwhich = level & 1;
368 	    list1 = which[iwhich];		/* area where runs are now */
369 	    list2 = which[++iwhich];		/* area for merged runs */
370 	    do {
371 		offset = stackp->offset;
372 		f1 = p1 = list1 + offset;		/* start of first run */
373 		p = tp2 = list2 + offset;	/* where merged run will go */
374 		t = NEXT(p);			/* where first run ends */
375 		f2 = l1 = POTHER(t, list2, list1); /* ... on the other side */
376 		t = NEXT(t);			/* where second runs ends */
377 		l2 = POTHER(t, list2, list1);	/* ... on the other side */
378 		offset = PNELEM(list2, t);
379 		while (f1 < l1 && f2 < l2) {
380 		    /* If head 1 is larger than head 2, find ALL the elements
381 		    ** in list 2 strictly less than head1, write them all,
382 		    ** then head 1.  Then compare the new heads, and repeat,
383 		    ** until one or both lists are exhausted.
384 		    **
385 		    ** In all comparisons (after establishing
386 		    ** which head to merge) the item to merge
387 		    ** (at pointer q) is the first operand of
388 		    ** the comparison.  When we want to know
389 		    ** if ``q is strictly less than the other'',
390 		    ** we can't just do
391 		    **    cmp(q, other) < 0
392 		    ** because stability demands that we treat equality
393 		    ** as high when q comes from l2, and as low when
394 		    ** q was from l1.  So we ask the question by doing
395 		    **    cmp(q, other) <= sense
396 		    ** and make sense == 0 when equality should look low,
397 		    ** and -1 when equality should look high.
398 		    */
399 
400 
401 		    if (cmp(aTHX_ *f1, *f2) <= 0) {
402 			q = f2; b = f1; t = l1;
403 			sense = -1;
404 		    } else {
405 			q = f1; b = f2; t = l2;
406 			sense = 0;
407 		    }
408 
409 
410 		    /* ramp up
411 		    **
412 		    ** Leave t at something strictly
413 		    ** greater than q (or at the end of the list),
414 		    ** and b at something strictly less than q.
415 		    */
416 		    for (i = 1, run = 0 ;;) {
417 			if ((p = PINDEX(b, i)) >= t) {
418 			    /* off the end */
419 			    if (((p = PINDEX(t, -1)) > b) &&
420 				(cmp(aTHX_ *q, *p) <= sense))
421 				 t = p;
422 			    else b = p;
423 			    break;
424 			} else if (cmp(aTHX_ *q, *p) <= sense) {
425 			    t = p;
426 			    break;
427 			} else b = p;
428 			if (++run >= RTHRESH) i += i;
429 		    }
430 
431 
432 		    /* q is known to follow b and must be inserted before t.
433 		    ** Increment b, so the range of possibilities is [b,t).
434 		    ** Round binary split down, to favor early appearance.
435 		    ** Adjust b and t until q belongs just before t.
436 		    */
437 
438 		    b++;
439 		    while (b < t) {
440 			p = PINDEX(b, (PNELEM(b, t) - 1) / 2);
441 			if (cmp(aTHX_ *q, *p) <= sense) {
442 			    t = p;
443 			} else b = p + 1;
444 		    }
445 
446 
447 		    /* Copy all the strictly low elements */
448 
449 		    if (q == f1) {
450 			FROMTOUPTO(f2, tp2, t);
451 			*tp2++ = *f1++;
452 		    } else {
453 			FROMTOUPTO(f1, tp2, t);
454 			*tp2++ = *f2++;
455 		    }
456 		}
457 
458 
459 		/* Run out remaining list */
460 		if (f1 == l1) {
461 		       if (f2 < l2) FROMTOUPTO(f2, tp2, l2);
462 		} else              FROMTOUPTO(f1, tp2, l1);
463 		p1 = NEXT(p1) = POTHER(tp2, list2, list1);
464 
465 		if (--level == 0) goto done;
466 		--stackp;
467 		t = list1; list1 = list2; list2 = t;	/* swap lists */
468 	    } while ((runs = stackp->runs) == 0);
469 	}
470 
471 
472 	stackp->runs = 0;		/* current run will finish level */
473 	/* While there are more than 2 runs remaining,
474 	 * turn them into exactly 2 runs (at the "other" level),
475 	 * each made up of approximately half the runs.
476 	 * Stack the second half for later processing,
477 	 * and set about producing the first half now.
478 	 */
479 	while (runs > 2) {
480 	    ++level;
481 	    ++stackp;
482 	    stackp->offset = offset;
483 	    runs -= stackp->runs = runs / 2;
484 	}
485 	/* We must construct a single run from 1 or 2 runs.
486 	 * All the original runs are in which[0] == base.
487 	 * The run we construct must end up in which[level&1].
488 	 */
489 	iwhich = level & 1;
490 	if (runs == 1) {
491 	    /* Constructing a single run from a single run.
492 	     * If it's where it belongs already, there's nothing to do.
493 	     * Otherwise, copy it to where it belongs.
494 	     * A run of 1 is either a singleton at level 0,
495 	     * or the second half of a split 3.  In neither event
496 	     * is it necessary to set offset.  It will be set by the merge
497 	     * that immediately follows.
498 	     */
499 	    if (iwhich) {	/* Belongs in aux, currently in base */
500 		f1 = b = PINDEX(base, offset);	/* where list starts */
501 		f2 = PINDEX(aux, offset);	/* where list goes */
502 		t = NEXT(f2);			/* where list will end */
503 		offset = PNELEM(aux, t);	/* offset thereof */
504 		t = PINDEX(base, offset);	/* where it currently ends */
505 		FROMTOUPTO(f1, f2, t);		/* copy */
506 		NEXT(b) = t;			/* set up parallel pointer */
507 	    } else if (level == 0) goto done;	/* single run at level 0 */
508 	} else {
509 	    /* Constructing a single run from two runs.
510 	     * The merge code at the top will do that.
511 	     * We need only make sure the two runs are in the "other" array,
512 	     * so they'll end up in the correct array after the merge.
513 	     */
514 	    ++level;
515 	    ++stackp;
516 	    stackp->offset = offset;
517 	    stackp->runs = 0;	/* take care of both runs, trigger merge */
518 	    if (!iwhich) {	/* Merged runs belong in aux, copy 1st */
519 		f1 = b = PINDEX(base, offset);	/* where first run starts */
520 		f2 = PINDEX(aux, offset);	/* where it will be copied */
521 		t = NEXT(f2);			/* where first run will end */
522 		offset = PNELEM(aux, t);	/* offset thereof */
523 		p = PINDEX(base, offset);	/* end of first run */
524 		t = NEXT(t);			/* where second run will end */
525 		t = PINDEX(base, PNELEM(aux, t)); /* where it now ends */
526 		FROMTOUPTO(f1, f2, t);		/* copy both runs */
527 		NEXT(b) = p;			/* paralled pointer for 1st */
528 		NEXT(p) = t;			/* ... and for second */
529 	    }
530 	}
531     }
532 done:
533     if (aux != small) Safefree(aux);	/* free iff allocated */
534     return;
535 }
536 
537 /*
538  * The quicksort implementation was derived from source code contributed
539  * by Tom Horsley.
540  *
541  * NOTE: this code was derived from Tom Horsley's qsort replacement
542  * and should not be confused with the original code.
543  */
544 
545 /* Copyright (C) Tom Horsley, 1997. All rights reserved.
546 
547    Permission granted to distribute under the same terms as perl which are
548    (briefly):
549 
550     This program is free software; you can redistribute it and/or modify
551     it under the terms of either:
552 
553 	a) the GNU General Public License as published by the Free
554 	Software Foundation; either version 1, or (at your option) any
555 	later version, or
556 
557 	b) the "Artistic License" which comes with this Kit.
558 
559    Details on the perl license can be found in the perl source code which
560    may be located via the www.perl.com web page.
561 
562    This is the most wonderfulest possible qsort I can come up with (and
563    still be mostly portable) My (limited) tests indicate it consistently
564    does about 20% fewer calls to compare than does the qsort in the Visual
565    C++ library, other vendors may vary.
566 
567    Some of the ideas in here can be found in "Algorithms" by Sedgewick,
568    others I invented myself (or more likely re-invented since they seemed
569    pretty obvious once I watched the algorithm operate for a while).
570 
571    Most of this code was written while watching the Marlins sweep the Giants
572    in the 1997 National League Playoffs - no Braves fans allowed to use this
573    code (just kidding :-).
574 
575    I realize that if I wanted to be true to the perl tradition, the only
576    comment in this file would be something like:
577 
578    ...they shuffled back towards the rear of the line. 'No, not at the
579    rear!'  the slave-driver shouted. 'Three files up. And stay there...
580 
581    However, I really needed to violate that tradition just so I could keep
582    track of what happens myself, not to mention some poor fool trying to
583    understand this years from now :-).
584 */
585 
586 /* ********************************************************** Configuration */
587 
588 #ifndef QSORT_ORDER_GUESS
589 #define QSORT_ORDER_GUESS 2	/* Select doubling version of the netBSD trick */
590 #endif
591 
592 /* QSORT_MAX_STACK is the largest number of partitions that can be stacked up for
593    future processing - a good max upper bound is log base 2 of memory size
594    (32 on 32 bit machines, 64 on 64 bit machines, etc). In reality can
595    safely be smaller than that since the program is taking up some space and
596    most operating systems only let you grab some subset of contiguous
597    memory (not to mention that you are normally sorting data larger than
598    1 byte element size :-).
599 */
600 #ifndef QSORT_MAX_STACK
601 #define QSORT_MAX_STACK 32
602 #endif
603 
604 /* QSORT_BREAK_EVEN is the size of the largest partition we should insertion sort.
605    Anything bigger and we use qsort. If you make this too small, the qsort
606    will probably break (or become less efficient), because it doesn't expect
607    the middle element of a partition to be the same as the right or left -
608    you have been warned).
609 */
610 #ifndef QSORT_BREAK_EVEN
611 #define QSORT_BREAK_EVEN 6
612 #endif
613 
614 /* QSORT_PLAY_SAFE is the size of the largest partition we're willing
615    to go quadratic on.  We innoculate larger partitions against
616    quadratic behavior by shuffling them before sorting.  This is not
617    an absolute guarantee of non-quadratic behavior, but it would take
618    staggeringly bad luck to pick extreme elements as the pivot
619    from randomized data.
620 */
621 #ifndef QSORT_PLAY_SAFE
622 #define QSORT_PLAY_SAFE 255
623 #endif
624 
625 /* ************************************************************* Data Types */
626 
627 /* hold left and right index values of a partition waiting to be sorted (the
628    partition includes both left and right - right is NOT one past the end or
629    anything like that).
630 */
631 struct partition_stack_entry {
632    int left;
633    int right;
634 #ifdef QSORT_ORDER_GUESS
635    int qsort_break_even;
636 #endif
637 };
638 
639 /* ******************************************************* Shorthand Macros */
640 
641 /* Note that these macros will be used from inside the qsort function where
642    we happen to know that the variable 'elt_size' contains the size of an
643    array element and the variable 'temp' points to enough space to hold a
644    temp element and the variable 'array' points to the array being sorted
645    and 'compare' is the pointer to the compare routine.
646 
647    Also note that there are very many highly architecture specific ways
648    these might be sped up, but this is simply the most generally portable
649    code I could think of.
650 */
651 
652 /* Return < 0 == 0 or > 0 as the value of elt1 is < elt2, == elt2, > elt2
653 */
654 #define qsort_cmp(elt1, elt2) \
655    ((*compare)(aTHX_ array[elt1], array[elt2]))
656 
657 #ifdef QSORT_ORDER_GUESS
658 #define QSORT_NOTICE_SWAP swapped++;
659 #else
660 #define QSORT_NOTICE_SWAP
661 #endif
662 
663 /* swaps contents of array elements elt1, elt2.
664 */
665 #define qsort_swap(elt1, elt2) \
666    STMT_START { \
667       QSORT_NOTICE_SWAP \
668       temp = array[elt1]; \
669       array[elt1] = array[elt2]; \
670       array[elt2] = temp; \
671    } STMT_END
672 
673 /* rotate contents of elt1, elt2, elt3 such that elt1 gets elt2, elt2 gets
674    elt3 and elt3 gets elt1.
675 */
676 #define qsort_rotate(elt1, elt2, elt3) \
677    STMT_START { \
678       QSORT_NOTICE_SWAP \
679       temp = array[elt1]; \
680       array[elt1] = array[elt2]; \
681       array[elt2] = array[elt3]; \
682       array[elt3] = temp; \
683    } STMT_END
684 
685 /* ************************************************************ Debug stuff */
686 
687 #ifdef QSORT_DEBUG
688 
689 static void
690 break_here()
691 {
692    return; /* good place to set a breakpoint */
693 }
694 
695 #define qsort_assert(t) (void)( (t) || (break_here(), 0) )
696 
697 static void
698 doqsort_all_asserts(
699    void * array,
700    size_t num_elts,
701    size_t elt_size,
702    int (*compare)(const void * elt1, const void * elt2),
703    int pc_left, int pc_right, int u_left, int u_right)
704 {
705    int i;
706 
707    qsort_assert(pc_left <= pc_right);
708    qsort_assert(u_right < pc_left);
709    qsort_assert(pc_right < u_left);
710    for (i = u_right + 1; i < pc_left; ++i) {
711       qsort_assert(qsort_cmp(i, pc_left) < 0);
712    }
713    for (i = pc_left; i < pc_right; ++i) {
714       qsort_assert(qsort_cmp(i, pc_right) == 0);
715    }
716    for (i = pc_right + 1; i < u_left; ++i) {
717       qsort_assert(qsort_cmp(pc_right, i) < 0);
718    }
719 }
720 
721 #define qsort_all_asserts(PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT) \
722    doqsort_all_asserts(array, num_elts, elt_size, compare, \
723                  PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT)
724 
725 #else
726 
727 #define qsort_assert(t) ((void)0)
728 
729 #define qsort_all_asserts(PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT) ((void)0)
730 
731 #endif
732 
733 /* ****************************************************************** qsort */
734 
735 STATIC void /* the standard unstable (u) quicksort (qsort) */
736 S_qsortsvu(pTHX_ SV ** array, size_t num_elts, SVCOMPARE_t compare)
737 {
738    register SV * temp;
739 
740    struct partition_stack_entry partition_stack[QSORT_MAX_STACK];
741    int next_stack_entry = 0;
742 
743    int part_left;
744    int part_right;
745 #ifdef QSORT_ORDER_GUESS
746    int qsort_break_even;
747    int swapped;
748 #endif
749 
750    /* Make sure we actually have work to do.
751    */
752    if (num_elts <= 1) {
753       return;
754    }
755 
756    /* Innoculate large partitions against quadratic behavior */
757    if (num_elts > QSORT_PLAY_SAFE) {
758       register size_t n, j;
759       register SV **q;
760       for (n = num_elts, q = array; n > 1; ) {
761          j = (size_t)(n-- * Drand01());
762          temp = q[j];
763          q[j] = q[n];
764          q[n] = temp;
765       }
766    }
767 
768    /* Setup the initial partition definition and fall into the sorting loop
769    */
770    part_left = 0;
771    part_right = (int)(num_elts - 1);
772 #ifdef QSORT_ORDER_GUESS
773    qsort_break_even = QSORT_BREAK_EVEN;
774 #else
775 #define qsort_break_even QSORT_BREAK_EVEN
776 #endif
777    for ( ; ; ) {
778       if ((part_right - part_left) >= qsort_break_even) {
779          /* OK, this is gonna get hairy, so lets try to document all the
780             concepts and abbreviations and variables and what they keep
781             track of:
782 
783             pc: pivot chunk - the set of array elements we accumulate in the
784                 middle of the partition, all equal in value to the original
785                 pivot element selected. The pc is defined by:
786 
787                 pc_left - the leftmost array index of the pc
788                 pc_right - the rightmost array index of the pc
789 
790                 we start with pc_left == pc_right and only one element
791                 in the pivot chunk (but it can grow during the scan).
792 
793             u:  uncompared elements - the set of elements in the partition
794                 we have not yet compared to the pivot value. There are two
795                 uncompared sets during the scan - one to the left of the pc
796                 and one to the right.
797 
798                 u_right - the rightmost index of the left side's uncompared set
799                 u_left - the leftmost index of the right side's uncompared set
800 
801                 The leftmost index of the left sides's uncompared set
802                 doesn't need its own variable because it is always defined
803                 by the leftmost edge of the whole partition (part_left). The
804                 same goes for the rightmost edge of the right partition
805                 (part_right).
806 
807                 We know there are no uncompared elements on the left once we
808                 get u_right < part_left and no uncompared elements on the
809                 right once u_left > part_right. When both these conditions
810                 are met, we have completed the scan of the partition.
811 
812                 Any elements which are between the pivot chunk and the
813                 uncompared elements should be less than the pivot value on
814                 the left side and greater than the pivot value on the right
815                 side (in fact, the goal of the whole algorithm is to arrange
816                 for that to be true and make the groups of less-than and
817                 greater-then elements into new partitions to sort again).
818 
819             As you marvel at the complexity of the code and wonder why it
820             has to be so confusing. Consider some of the things this level
821             of confusion brings:
822 
823             Once I do a compare, I squeeze every ounce of juice out of it. I
824             never do compare calls I don't have to do, and I certainly never
825             do redundant calls.
826 
827             I also never swap any elements unless I can prove there is a
828             good reason. Many sort algorithms will swap a known value with
829             an uncompared value just to get things in the right place (or
830             avoid complexity :-), but that uncompared value, once it gets
831             compared, may then have to be swapped again. A lot of the
832             complexity of this code is due to the fact that it never swaps
833             anything except compared values, and it only swaps them when the
834             compare shows they are out of position.
835          */
836          int pc_left, pc_right;
837          int u_right, u_left;
838 
839          int s;
840 
841          pc_left = ((part_left + part_right) / 2);
842          pc_right = pc_left;
843          u_right = pc_left - 1;
844          u_left = pc_right + 1;
845 
846          /* Qsort works best when the pivot value is also the median value
847             in the partition (unfortunately you can't find the median value
848             without first sorting :-), so to give the algorithm a helping
849             hand, we pick 3 elements and sort them and use the median value
850             of that tiny set as the pivot value.
851 
852             Some versions of qsort like to use the left middle and right as
853             the 3 elements to sort so they can insure the ends of the
854             partition will contain values which will stop the scan in the
855             compare loop, but when you have to call an arbitrarily complex
856             routine to do a compare, its really better to just keep track of
857             array index values to know when you hit the edge of the
858             partition and avoid the extra compare. An even better reason to
859             avoid using a compare call is the fact that you can drop off the
860             edge of the array if someone foolishly provides you with an
861             unstable compare function that doesn't always provide consistent
862             results.
863 
864             So, since it is simpler for us to compare the three adjacent
865             elements in the middle of the partition, those are the ones we
866             pick here (conveniently pointed at by u_right, pc_left, and
867             u_left). The values of the left, center, and right elements
868             are refered to as l c and r in the following comments.
869          */
870 
871 #ifdef QSORT_ORDER_GUESS
872          swapped = 0;
873 #endif
874          s = qsort_cmp(u_right, pc_left);
875          if (s < 0) {
876             /* l < c */
877             s = qsort_cmp(pc_left, u_left);
878             /* if l < c, c < r - already in order - nothing to do */
879             if (s == 0) {
880                /* l < c, c == r - already in order, pc grows */
881                ++pc_right;
882                qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
883             } else if (s > 0) {
884                /* l < c, c > r - need to know more */
885                s = qsort_cmp(u_right, u_left);
886                if (s < 0) {
887                   /* l < c, c > r, l < r - swap c & r to get ordered */
888                   qsort_swap(pc_left, u_left);
889                   qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
890                } else if (s == 0) {
891                   /* l < c, c > r, l == r - swap c&r, grow pc */
892                   qsort_swap(pc_left, u_left);
893                   --pc_left;
894                   qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
895                } else {
896                   /* l < c, c > r, l > r - make lcr into rlc to get ordered */
897                   qsort_rotate(pc_left, u_right, u_left);
898                   qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
899                }
900             }
901          } else if (s == 0) {
902             /* l == c */
903             s = qsort_cmp(pc_left, u_left);
904             if (s < 0) {
905                /* l == c, c < r - already in order, grow pc */
906                --pc_left;
907                qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
908             } else if (s == 0) {
909                /* l == c, c == r - already in order, grow pc both ways */
910                --pc_left;
911                ++pc_right;
912                qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
913             } else {
914                /* l == c, c > r - swap l & r, grow pc */
915                qsort_swap(u_right, u_left);
916                ++pc_right;
917                qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
918             }
919          } else {
920             /* l > c */
921             s = qsort_cmp(pc_left, u_left);
922             if (s < 0) {
923                /* l > c, c < r - need to know more */
924                s = qsort_cmp(u_right, u_left);
925                if (s < 0) {
926                   /* l > c, c < r, l < r - swap l & c to get ordered */
927                   qsort_swap(u_right, pc_left);
928                   qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
929                } else if (s == 0) {
930                   /* l > c, c < r, l == r - swap l & c, grow pc */
931                   qsort_swap(u_right, pc_left);
932                   ++pc_right;
933                   qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
934                } else {
935                   /* l > c, c < r, l > r - rotate lcr into crl to order */
936                   qsort_rotate(u_right, pc_left, u_left);
937                   qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
938                }
939             } else if (s == 0) {
940                /* l > c, c == r - swap ends, grow pc */
941                qsort_swap(u_right, u_left);
942                --pc_left;
943                qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
944             } else {
945                /* l > c, c > r - swap ends to get in order */
946                qsort_swap(u_right, u_left);
947                qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
948             }
949          }
950          /* We now know the 3 middle elements have been compared and
951             arranged in the desired order, so we can shrink the uncompared
952             sets on both sides
953          */
954          --u_right;
955          ++u_left;
956          qsort_all_asserts(pc_left, pc_right, u_left, u_right);
957 
958          /* The above massive nested if was the simple part :-). We now have
959             the middle 3 elements ordered and we need to scan through the
960             uncompared sets on either side, swapping elements that are on
961             the wrong side or simply shuffling equal elements around to get
962             all equal elements into the pivot chunk.
963          */
964 
965          for ( ; ; ) {
966             int still_work_on_left;
967             int still_work_on_right;
968 
969             /* Scan the uncompared values on the left. If I find a value
970                equal to the pivot value, move it over so it is adjacent to
971                the pivot chunk and expand the pivot chunk. If I find a value
972                less than the pivot value, then just leave it - its already
973                on the correct side of the partition. If I find a greater
974                value, then stop the scan.
975             */
976             while ((still_work_on_left = (u_right >= part_left))) {
977                s = qsort_cmp(u_right, pc_left);
978                if (s < 0) {
979                   --u_right;
980                } else if (s == 0) {
981                   --pc_left;
982                   if (pc_left != u_right) {
983                      qsort_swap(u_right, pc_left);
984                   }
985                   --u_right;
986                } else {
987                   break;
988                }
989                qsort_assert(u_right < pc_left);
990                qsort_assert(pc_left <= pc_right);
991                qsort_assert(qsort_cmp(u_right + 1, pc_left) <= 0);
992                qsort_assert(qsort_cmp(pc_left, pc_right) == 0);
993             }
994 
995             /* Do a mirror image scan of uncompared values on the right
996             */
997             while ((still_work_on_right = (u_left <= part_right))) {
998                s = qsort_cmp(pc_right, u_left);
999                if (s < 0) {
1000                   ++u_left;
1001                } else if (s == 0) {
1002                   ++pc_right;
1003                   if (pc_right != u_left) {
1004                      qsort_swap(pc_right, u_left);
1005                   }
1006                   ++u_left;
1007                } else {
1008                   break;
1009                }
1010                qsort_assert(u_left > pc_right);
1011                qsort_assert(pc_left <= pc_right);
1012                qsort_assert(qsort_cmp(pc_right, u_left - 1) <= 0);
1013                qsort_assert(qsort_cmp(pc_left, pc_right) == 0);
1014             }
1015 
1016             if (still_work_on_left) {
1017                /* I know I have a value on the left side which needs to be
1018                   on the right side, but I need to know more to decide
1019                   exactly the best thing to do with it.
1020                */
1021                if (still_work_on_right) {
1022                   /* I know I have values on both side which are out of
1023                      position. This is a big win because I kill two birds
1024                      with one swap (so to speak). I can advance the
1025                      uncompared pointers on both sides after swapping both
1026                      of them into the right place.
1027                   */
1028                   qsort_swap(u_right, u_left);
1029                   --u_right;
1030                   ++u_left;
1031                   qsort_all_asserts(pc_left, pc_right, u_left, u_right);
1032                } else {
1033                   /* I have an out of position value on the left, but the
1034                      right is fully scanned, so I "slide" the pivot chunk
1035                      and any less-than values left one to make room for the
1036                      greater value over on the right. If the out of position
1037                      value is immediately adjacent to the pivot chunk (there
1038                      are no less-than values), I can do that with a swap,
1039                      otherwise, I have to rotate one of the less than values
1040                      into the former position of the out of position value
1041                      and the right end of the pivot chunk into the left end
1042                      (got all that?).
1043                   */
1044                   --pc_left;
1045                   if (pc_left == u_right) {
1046                      qsort_swap(u_right, pc_right);
1047                      qsort_all_asserts(pc_left, pc_right-1, u_left, u_right-1);
1048                   } else {
1049                      qsort_rotate(u_right, pc_left, pc_right);
1050                      qsort_all_asserts(pc_left, pc_right-1, u_left, u_right-1);
1051                   }
1052                   --pc_right;
1053                   --u_right;
1054                }
1055             } else if (still_work_on_right) {
1056                /* Mirror image of complex case above: I have an out of
1057                   position value on the right, but the left is fully
1058                   scanned, so I need to shuffle things around to make room
1059                   for the right value on the left.
1060                */
1061                ++pc_right;
1062                if (pc_right == u_left) {
1063                   qsort_swap(u_left, pc_left);
1064                   qsort_all_asserts(pc_left+1, pc_right, u_left+1, u_right);
1065                } else {
1066                   qsort_rotate(pc_right, pc_left, u_left);
1067                   qsort_all_asserts(pc_left+1, pc_right, u_left+1, u_right);
1068                }
1069                ++pc_left;
1070                ++u_left;
1071             } else {
1072                /* No more scanning required on either side of partition,
1073                   break out of loop and figure out next set of partitions
1074                */
1075                break;
1076             }
1077          }
1078 
1079          /* The elements in the pivot chunk are now in the right place. They
1080             will never move or be compared again. All I have to do is decide
1081             what to do with the stuff to the left and right of the pivot
1082             chunk.
1083 
1084             Notes on the QSORT_ORDER_GUESS ifdef code:
1085 
1086             1. If I just built these partitions without swapping any (or
1087                very many) elements, there is a chance that the elements are
1088                already ordered properly (being properly ordered will
1089                certainly result in no swapping, but the converse can't be
1090                proved :-).
1091 
1092             2. A (properly written) insertion sort will run faster on
1093                already ordered data than qsort will.
1094 
1095             3. Perhaps there is some way to make a good guess about
1096                switching to an insertion sort earlier than partition size 6
1097                (for instance - we could save the partition size on the stack
1098                and increase the size each time we find we didn't swap, thus
1099                switching to insertion sort earlier for partitions with a
1100                history of not swapping).
1101 
1102             4. Naturally, if I just switch right away, it will make
1103                artificial benchmarks with pure ascending (or descending)
1104                data look really good, but is that a good reason in general?
1105                Hard to say...
1106          */
1107 
1108 #ifdef QSORT_ORDER_GUESS
1109          if (swapped < 3) {
1110 #if QSORT_ORDER_GUESS == 1
1111             qsort_break_even = (part_right - part_left) + 1;
1112 #endif
1113 #if QSORT_ORDER_GUESS == 2
1114             qsort_break_even *= 2;
1115 #endif
1116 #if QSORT_ORDER_GUESS == 3
1117             int prev_break = qsort_break_even;
1118             qsort_break_even *= qsort_break_even;
1119             if (qsort_break_even < prev_break) {
1120                qsort_break_even = (part_right - part_left) + 1;
1121             }
1122 #endif
1123          } else {
1124             qsort_break_even = QSORT_BREAK_EVEN;
1125          }
1126 #endif
1127 
1128          if (part_left < pc_left) {
1129             /* There are elements on the left which need more processing.
1130                Check the right as well before deciding what to do.
1131             */
1132             if (pc_right < part_right) {
1133                /* We have two partitions to be sorted. Stack the biggest one
1134                   and process the smallest one on the next iteration. This
1135                   minimizes the stack height by insuring that any additional
1136                   stack entries must come from the smallest partition which
1137                   (because it is smallest) will have the fewest
1138                   opportunities to generate additional stack entries.
1139                */
1140                if ((part_right - pc_right) > (pc_left - part_left)) {
1141                   /* stack the right partition, process the left */
1142                   partition_stack[next_stack_entry].left = pc_right + 1;
1143                   partition_stack[next_stack_entry].right = part_right;
1144 #ifdef QSORT_ORDER_GUESS
1145                   partition_stack[next_stack_entry].qsort_break_even = qsort_break_even;
1146 #endif
1147                   part_right = pc_left - 1;
1148                } else {
1149                   /* stack the left partition, process the right */
1150                   partition_stack[next_stack_entry].left = part_left;
1151                   partition_stack[next_stack_entry].right = pc_left - 1;
1152 #ifdef QSORT_ORDER_GUESS
1153                   partition_stack[next_stack_entry].qsort_break_even = qsort_break_even;
1154 #endif
1155                   part_left = pc_right + 1;
1156                }
1157                qsort_assert(next_stack_entry < QSORT_MAX_STACK);
1158                ++next_stack_entry;
1159             } else {
1160                /* The elements on the left are the only remaining elements
1161                   that need sorting, arrange for them to be processed as the
1162                   next partition.
1163                */
1164                part_right = pc_left - 1;
1165             }
1166          } else if (pc_right < part_right) {
1167             /* There is only one chunk on the right to be sorted, make it
1168                the new partition and loop back around.
1169             */
1170             part_left = pc_right + 1;
1171          } else {
1172             /* This whole partition wound up in the pivot chunk, so
1173                we need to get a new partition off the stack.
1174             */
1175             if (next_stack_entry == 0) {
1176                /* the stack is empty - we are done */
1177                break;
1178             }
1179             --next_stack_entry;
1180             part_left = partition_stack[next_stack_entry].left;
1181             part_right = partition_stack[next_stack_entry].right;
1182 #ifdef QSORT_ORDER_GUESS
1183             qsort_break_even = partition_stack[next_stack_entry].qsort_break_even;
1184 #endif
1185          }
1186       } else {
1187          /* This partition is too small to fool with qsort complexity, just
1188             do an ordinary insertion sort to minimize overhead.
1189          */
1190          int i;
1191          /* Assume 1st element is in right place already, and start checking
1192             at 2nd element to see where it should be inserted.
1193          */
1194          for (i = part_left + 1; i <= part_right; ++i) {
1195             int j;
1196             /* Scan (backwards - just in case 'i' is already in right place)
1197                through the elements already sorted to see if the ith element
1198                belongs ahead of one of them.
1199             */
1200             for (j = i - 1; j >= part_left; --j) {
1201                if (qsort_cmp(i, j) >= 0) {
1202                   /* i belongs right after j
1203                   */
1204                   break;
1205                }
1206             }
1207             ++j;
1208             if (j != i) {
1209                /* Looks like we really need to move some things
1210                */
1211 	       int k;
1212 	       temp = array[i];
1213 	       for (k = i - 1; k >= j; --k)
1214 		  array[k + 1] = array[k];
1215                array[j] = temp;
1216             }
1217          }
1218 
1219          /* That partition is now sorted, grab the next one, or get out
1220             of the loop if there aren't any more.
1221          */
1222 
1223          if (next_stack_entry == 0) {
1224             /* the stack is empty - we are done */
1225             break;
1226          }
1227          --next_stack_entry;
1228          part_left = partition_stack[next_stack_entry].left;
1229          part_right = partition_stack[next_stack_entry].right;
1230 #ifdef QSORT_ORDER_GUESS
1231          qsort_break_even = partition_stack[next_stack_entry].qsort_break_even;
1232 #endif
1233       }
1234    }
1235 
1236    /* Believe it or not, the array is sorted at this point! */
1237 }
1238 
1239 /* Stabilize what is, presumably, an otherwise unstable sort method.
1240  * We do that by allocating (or having on hand) an array of pointers
1241  * that is the same size as the original array of elements to be sorted.
1242  * We initialize this parallel array with the addresses of the original
1243  * array elements.  This indirection can make you crazy.
1244  * Some pictures can help.  After initializing, we have
1245  *
1246  *  indir                  list1
1247  * +----+                 +----+
1248  * |    | --------------> |    | ------> first element to be sorted
1249  * +----+                 +----+
1250  * |    | --------------> |    | ------> second element to be sorted
1251  * +----+                 +----+
1252  * |    | --------------> |    | ------> third element to be sorted
1253  * +----+                 +----+
1254  *  ...
1255  * +----+                 +----+
1256  * |    | --------------> |    | ------> n-1st element to be sorted
1257  * +----+                 +----+
1258  * |    | --------------> |    | ------> n-th element to be sorted
1259  * +----+                 +----+
1260  *
1261  * During the sort phase, we leave the elements of list1 where they are,
1262  * and sort the pointers in the indirect array in the same order determined
1263  * by the original comparison routine on the elements pointed to.
1264  * Because we don't move the elements of list1 around through
1265  * this phase, we can break ties on elements that compare equal
1266  * using their address in the list1 array, ensuring stabilty.
1267  * This leaves us with something looking like
1268  *
1269  *  indir                  list1
1270  * +----+                 +----+
1271  * |    | --+       +---> |    | ------> first element to be sorted
1272  * +----+   |       |     +----+
1273  * |    | --|-------|---> |    | ------> second element to be sorted
1274  * +----+   |       |     +----+
1275  * |    | --|-------+ +-> |    | ------> third element to be sorted
1276  * +----+   |         |   +----+
1277  *  ...
1278  * +----+    | |   | |    +----+
1279  * |    | ---|-+   | +--> |    | ------> n-1st element to be sorted
1280  * +----+    |     |      +----+
1281  * |    | ---+     +----> |    | ------> n-th element to be sorted
1282  * +----+                 +----+
1283  *
1284  * where the i-th element of the indirect array points to the element
1285  * that should be i-th in the sorted array.  After the sort phase,
1286  * we have to put the elements of list1 into the places
1287  * dictated by the indirect array.
1288  */
1289 
1290 
1291 static I32
1292 cmpindir(pTHX_ gptr a, gptr b)
1293 {
1294     I32 sense;
1295     gptr *ap = (gptr *)a;
1296     gptr *bp = (gptr *)b;
1297 
1298     if ((sense = PL_sort_RealCmp(aTHX_ *ap, *bp)) == 0)
1299 	 sense = (ap > bp) ? 1 : ((ap < bp) ? -1 : 0);
1300     return sense;
1301 }
1302 
1303 STATIC void
1304 S_qsortsv(pTHX_ gptr *list1, size_t nmemb, SVCOMPARE_t cmp)
1305 {
1306     SV *hintsv;
1307 
1308     if (SORTHINTS(hintsv) & HINT_SORT_STABLE) {
1309 	 register gptr **pp, *q;
1310 	 register size_t n, j, i;
1311 	 gptr *small[SMALLSORT], **indir, tmp;
1312 	 SVCOMPARE_t savecmp;
1313 	 if (nmemb <= 1) return;     /* sorted trivially */
1314 
1315 	 /* Small arrays can use the stack, big ones must be allocated */
1316 	 if (nmemb <= SMALLSORT) indir = small;
1317 	 else { New(1799, indir, nmemb, gptr *); }
1318 
1319 	 /* Copy pointers to original array elements into indirect array */
1320 	 for (n = nmemb, pp = indir, q = list1; n--; ) *pp++ = q++;
1321 
1322 	 savecmp = PL_sort_RealCmp;	/* Save current comparison routine, if any */
1323 	 PL_sort_RealCmp = cmp;	/* Put comparison routine where cmpindir can find it */
1324 
1325 	 /* sort, with indirection */
1326 	 S_qsortsvu(aTHX_ (gptr *)indir, nmemb, cmpindir);
1327 
1328 	 pp = indir;
1329 	 q = list1;
1330 	 for (n = nmemb; n--; ) {
1331 	      /* Assert A: all elements of q with index > n are already
1332 	       * in place.  This is vacuosly true at the start, and we
1333 	       * put element n where it belongs below (if it wasn't
1334 	       * already where it belonged). Assert B: we only move
1335 	       * elements that aren't where they belong,
1336 	       * so, by A, we never tamper with elements above n.
1337 	       */
1338 	      j = pp[n] - q;		/* This sets j so that q[j] is
1339 					 * at pp[n].  *pp[j] belongs in
1340 					 * q[j], by construction.
1341 					 */
1342 	      if (n != j) {		/* all's well if n == j */
1343 		   tmp = q[j];		/* save what's in q[j] */
1344 		   do {
1345 			q[j] = *pp[j];	/* put *pp[j] where it belongs */
1346 			i = pp[j] - q;	/* the index in q of the element
1347 					 * just moved */
1348 			pp[j] = q + j;	/* this is ok now */
1349 		   } while ((j = i) != n);
1350 		   /* There are only finitely many (nmemb) addresses
1351 		    * in the pp array.
1352 		    * So we must eventually revisit an index we saw before.
1353 		    * Suppose the first revisited index is k != n.
1354 		    * An index is visited because something else belongs there.
1355 		    * If we visit k twice, then two different elements must
1356 		    * belong in the same place, which cannot be.
1357 		    * So j must get back to n, the loop terminates,
1358 		    * and we put the saved element where it belongs.
1359 		    */
1360 		   q[n] = tmp;		/* put what belongs into
1361 					 * the n-th element */
1362 	      }
1363 	 }
1364 
1365 	/* free iff allocated */
1366 	 if (indir != small) { Safefree(indir); }
1367 	 /* restore prevailing comparison routine */
1368 	 PL_sort_RealCmp = savecmp;
1369     } else {
1370 	 S_qsortsvu(aTHX_ list1, nmemb, cmp);
1371     }
1372 }
1373 
1374 /*
1375 =head1 Array Manipulation Functions
1376 
1377 =for apidoc sortsv
1378 
1379 Sort an array. Here is an example:
1380 
1381     sortsv(AvARRAY(av), av_len(av)+1, Perl_sv_cmp_locale);
1382 
1383 See lib/sort.pm for details about controlling the sorting algorithm.
1384 
1385 =cut
1386 */
1387 
1388 void
1389 Perl_sortsv(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp)
1390 {
1391     void (*sortsvp)(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp) =
1392         S_mergesortsv;
1393     SV *hintsv;
1394     I32 hints;
1395 
1396     /*  Sun's Compiler (cc: WorkShop Compilers 4.2 30 Oct 1996 C 4.2) used
1397 	to miscompile this function under optimization -O.  If you get test
1398 	errors related to picking the correct sort() function, try recompiling
1399 	this file without optimiziation.  -- A.D.  4/2002.
1400     */
1401     hints = SORTHINTS(hintsv);
1402     if (hints & HINT_SORT_QUICKSORT) {
1403 	sortsvp = S_qsortsv;
1404     }
1405     else {
1406 	/* The default as of 5.8.0 is mergesort */
1407 	sortsvp = S_mergesortsv;
1408     }
1409 
1410     sortsvp(aTHX_ array, nmemb, cmp);
1411 }
1412 
1413 PP(pp_sort)
1414 {
1415     dSP; dMARK; dORIGMARK;
1416     register SV **up;
1417     SV **myorigmark = ORIGMARK;
1418     register I32 max;
1419     HV *stash;
1420     GV *gv;
1421     CV *cv = 0;
1422     I32 gimme = GIMME;
1423     OP* nextop = PL_op->op_next;
1424     I32 overloading = 0;
1425     bool hasargs = FALSE;
1426     I32 is_xsub = 0;
1427 
1428     if (gimme != G_ARRAY) {
1429 	SP = MARK;
1430 	RETPUSHUNDEF;
1431     }
1432 
1433     ENTER;
1434     SAVEVPTR(PL_sortcop);
1435     if (PL_op->op_flags & OPf_STACKED) {
1436 	if (PL_op->op_flags & OPf_SPECIAL) {
1437 	    OP *kid = cLISTOP->op_first->op_sibling;	/* pass pushmark */
1438 	    kid = kUNOP->op_first;			/* pass rv2gv */
1439 	    kid = kUNOP->op_first;			/* pass leave */
1440 	    PL_sortcop = kid->op_next;
1441 	    stash = CopSTASH(PL_curcop);
1442 	}
1443 	else {
1444 	    cv = sv_2cv(*++MARK, &stash, &gv, 0);
1445 	    if (cv && SvPOK(cv)) {
1446 		STRLEN n_a;
1447 		char *proto = SvPV((SV*)cv, n_a);
1448 		if (proto && strEQ(proto, "$$")) {
1449 		    hasargs = TRUE;
1450 		}
1451 	    }
1452 	    if (!(cv && CvROOT(cv))) {
1453 		if (cv && CvXSUB(cv)) {
1454 		    is_xsub = 1;
1455 		}
1456 		else if (gv) {
1457 		    SV *tmpstr = sv_newmortal();
1458 		    gv_efullname3(tmpstr, gv, Nullch);
1459 		    DIE(aTHX_ "Undefined sort subroutine \"%"SVf"\" called",
1460 			tmpstr);
1461 		}
1462 		else {
1463 		    DIE(aTHX_ "Undefined subroutine in sort");
1464 		}
1465 	    }
1466 
1467 	    if (is_xsub)
1468 		PL_sortcop = (OP*)cv;
1469 	    else {
1470 		PL_sortcop = CvSTART(cv);
1471 		SAVEVPTR(CvROOT(cv)->op_ppaddr);
1472 		CvROOT(cv)->op_ppaddr = PL_ppaddr[OP_NULL];
1473 
1474 		PAD_SET_CUR(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*)PAD_SVl(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_CURPAD_SAVE(cx->blk_sub);
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*)PAD_SVl(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