Project: Perl
Code Location: git://perl5.git.perl.org/perl.gitmaint-5.12
Browse
/
Outline
Download File
pp_sort.c
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
/*    pp_sort.c
 *
 *    Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000,
 *    2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008 by Larry Wall and others
 *
 *    You may distribute under the terms of either the GNU General Public
 *    License or the Artistic License, as specified in the README file.
 *
 */

/*
 *   ...they shuffled back towards the rear of the line.  'No, not at the
 *   rear!' the slave-driver shouted.  'Three files up. And stay there...
 *
 *     [p.931 of _The Lord of the Rings_, VI/ii: "The Land of Shadow"]
 */

/* This file contains pp ("push/pop") functions that
 * execute the opcodes that make up a perl program. A typical pp function
 * expects to find its arguments on the stack, and usually pushes its
 * results onto the stack, hence the 'pp' terminology. Each OP structure
 * contains a pointer to the relevant pp_foo() function.
 *
 * This particular file just contains pp_sort(), which is complex
 * enough to merit its own file! See the other pp*.c files for the rest of
 * the pp_ functions.
 */

#include "EXTERN.h"
#define PERL_IN_PP_SORT_C
#include "perl.h"

#if defined(UNDER_CE)
/* looks like 'small' is reserved word for WINCE (or somesuch)*/
#define	small xsmall
#endif

#define sv_cmp_static Perl_sv_cmp
#define sv_cmp_locale_static Perl_sv_cmp_locale

#ifndef SMALLSORT
#define	SMALLSORT (200)
#endif

/* Flags for qsortsv and mergesortsv */
#define SORTf_DESC   1
#define SORTf_STABLE 2
#define SORTf_QSORT  4

/*
 * The mergesort implementation is by Peter M. Mcilroy <pmcilroy@lucent.com>.
 *
 * The original code was written in conjunction with BSD Computer Software
 * Research Group at University of California, Berkeley.
 *
 * See also: "Optimistic Merge Sort" (SODA '92)
 *
 * The integration to Perl is by John P. Linderman <jpl@research.att.com>.
 *
 * The code can be distributed under the same terms as Perl itself.
 *
 */


typedef char * aptr;		/* pointer for arithmetic on sizes */
typedef SV * gptr;		/* pointers in our lists */

/* Binary merge internal sort, with a few special mods
** for the special perl environment it now finds itself in.
**
** Things that were once options have been hotwired
** to values suitable for this use.  In particular, we'll always
** initialize looking for natural runs, we'll always produce stable
** output, and we'll always do Peter McIlroy's binary merge.
*/

/* Pointer types for arithmetic and storage and convenience casts */

#define	APTR(P)	((aptr)(P))
#define	GPTP(P)	((gptr *)(P))
#define GPPP(P) ((gptr **)(P))


/* byte offset from pointer P to (larger) pointer Q */
#define	BYTEOFF(P, Q) (APTR(Q) - APTR(P))

#define PSIZE sizeof(gptr)

/* If PSIZE is power of 2, make PSHIFT that power, if that helps */

#ifdef	PSHIFT
#define	PNELEM(P, Q)	(BYTEOFF(P,Q) >> (PSHIFT))
#define	PNBYTE(N)	((N) << (PSHIFT))
#define	PINDEX(P, N)	(GPTP(APTR(P) + PNBYTE(N)))
#else
/* Leave optimization to compiler */
#define	PNELEM(P, Q)	(GPTP(Q) - GPTP(P))
#define	PNBYTE(N)	((N) * (PSIZE))
#define	PINDEX(P, N)	(GPTP(P) + (N))
#endif

/* Pointer into other corresponding to pointer into this */
#define	POTHER(P, THIS, OTHER) GPTP(APTR(OTHER) + BYTEOFF(THIS,P))

#define FROMTOUPTO(src, dst, lim) do *dst++ = *src++; while(src<lim)


/* Runs are identified by a pointer in the auxilliary list.
** The pointer is at the start of the list,
** and it points to the start of the next list.
** NEXT is used as an lvalue, too.
*/

#define	NEXT(P)		(*GPPP(P))


/* PTHRESH is the minimum number of pairs with the same sense to justify
** checking for a run and extending it.  Note that PTHRESH counts PAIRS,
** not just elements, so PTHRESH == 8 means a run of 16.
*/

#define	PTHRESH (8)

/* RTHRESH is the number of elements in a run that must compare low
** to the low element from the opposing run before we justify
** doing a binary rampup instead of single stepping.
** In random input, N in a row low should only happen with
** probability 2^(1-N), so we can risk that we are dealing
** with orderly input without paying much when we aren't.
*/

#define RTHRESH (6)


/*
** Overview of algorithm and variables.
** The array of elements at list1 will be organized into runs of length 2,
** or runs of length >= 2 * PTHRESH.  We only try to form long runs when
** PTHRESH adjacent pairs compare in the same way, suggesting overall order.
**
** Unless otherwise specified, pair pointers address the first of two elements.
**
** b and b+1 are a pair that compare with sense "sense".
** b is the "bottom" of adjacent pairs that might form a longer run.
**
** p2 parallels b in the list2 array, where runs are defined by
** a pointer chain.
**
** t represents the "top" of the adjacent pairs that might extend
** the run beginning at b.  Usually, t addresses a pair
** that compares with opposite sense from (b,b+1).
** However, it may also address a singleton element at the end of list1,
** or it may be equal to "last", the first element beyond list1.
**
** r addresses the Nth pair following b.  If this would be beyond t,
** we back it off to t.  Only when r is less than t do we consider the
** run long enough to consider checking.
**
** q addresses a pair such that the pairs at b through q already form a run.
** Often, q will equal b, indicating we only are sure of the pair itself.
** However, a search on the previous cycle may have revealed a longer run,
** so q may be greater than b.
**
** p is used to work back from a candidate r, trying to reach q,
** which would mean b through r would be a run.  If we discover such a run,
** we start q at r and try to push it further towards t.
** If b through r is NOT a run, we detect the wrong order at (p-1,p).
** In any event, after the check (if any), we have two main cases.
**
** 1) Short run.  b <= q < p <= r <= t.
**	b through q is a run (perhaps trivial)
**	q through p are uninteresting pairs
**	p through r is a run
**
** 2) Long run.  b < r <= q < t.
**	b through q is a run (of length >= 2 * PTHRESH)
**
** Note that degenerate cases are not only possible, but likely.
** For example, if the pair following b compares with opposite sense,
** then b == q < p == r == t.
*/


static IV
dynprep(pTHX_ gptr *list1, gptr *list2, size_t nmemb, const SVCOMPARE_t cmp)
{
    I32 sense;
    register gptr *b, *p, *q, *t, *p2;
    register gptr *last, *r;
    IV runs = 0;

    b = list1;
    last = PINDEX(b, nmemb);
    sense = (cmp(aTHX_ *b, *(b+1)) > 0);
    for (p2 = list2; b < last; ) {
	/* We just started, or just reversed sense.
	** Set t at end of pairs with the prevailing sense.
	*/
	for (p = b+2, t = p; ++p < last; t = ++p) {
	    if ((cmp(aTHX_ *t, *p) > 0) != sense) break;
	}
	q = b;
	/* Having laid out the playing field, look for long runs */
	do {
	    p = r = b + (2 * PTHRESH);
	    if (r >= t) p = r = t;	/* too short to care about */
	    else {
		while (((cmp(aTHX_ *(p-1), *p) > 0) == sense) &&
		       ((p -= 2) > q)) {}
		if (p <= q) {
		    /* b through r is a (long) run.
		    ** Extend it as far as possible.
		    */
		    p = q = r;
		    while (((p += 2) < t) &&
			   ((cmp(aTHX_ *(p-1), *p) > 0) == sense)) q = p;
		    r = p = q + 2;	/* no simple pairs, no after-run */
		}
	    }
	    if (q > b) {		/* run of greater than 2 at b */
		gptr *savep = p;

		p = q += 2;
		/* pick up singleton, if possible */
		if ((p == t) &&
		    ((t + 1) == last) &&
		    ((cmp(aTHX_ *(p-1), *p) > 0) == sense))
		    savep = r = p = q = last;
		p2 = NEXT(p2) = p2 + (p - b); ++runs;
		if (sense)
		    while (b < --p) {
			const gptr c = *b;
			*b++ = *p;
			*p = c;
		    }
		p = savep;
	    }
	    while (q < p) {		/* simple pairs */
		p2 = NEXT(p2) = p2 + 2; ++runs;
		if (sense) {
		    const gptr c = *q++;
		    *(q-1) = *q;
		    *q++ = c;
		} else q += 2;
	    }
	    if (((b = p) == t) && ((t+1) == last)) {
		NEXT(p2) = p2 + 1; ++runs;
		b++;
	    }
	    q = r;
	} while (b < t);
	sense = !sense;
    }
    return runs;
}


/* The original merge sort, in use since 5.7, was as fast as, or faster than,
 * qsort on many platforms, but slower than qsort, conspicuously so,
 * on others.  The most likely explanation was platform-specific
 * differences in cache sizes and relative speeds.
 *
 * The quicksort divide-and-conquer algorithm guarantees that, as the
 * problem is subdivided into smaller and smaller parts, the parts
 * fit into smaller (and faster) caches.  So it doesn't matter how
 * many levels of cache exist, quicksort will "find" them, and,
 * as long as smaller is faster, take advantage of them.
 *
 * By contrast, consider how the original mergesort algorithm worked.
 * Suppose we have five runs (each typically of length 2 after dynprep).
 * 
 * pass               base                        aux
 *  0              1 2 3 4 5
 *  1                                           12 34 5
 *  2                1234 5
 *  3                                            12345
 *  4                 12345
 *
 * Adjacent pairs are merged in "grand sweeps" through the input.
 * This means, on pass 1, the records in runs 1 and 2 aren't revisited until
 * runs 3 and 4 are merged and the runs from run 5 have been copied.
 * The only cache that matters is one large enough to hold *all* the input.
 * On some platforms, this may be many times slower than smaller caches.
 *
 * The following pseudo-code uses the same basic merge algorithm,
 * but in a divide-and-conquer way.
 *
 * # merge $runs runs at offset $offset of list $list1 into $list2.
 * # all unmerged runs ($runs == 1) originate in list $base.
 * sub mgsort2 {
 *     my ($offset, $runs, $base, $list1, $list2) = @_;
 *
 *     if ($runs == 1) {
 *         if ($list1 is $base) copy run to $list2
 *         return offset of end of list (or copy)
 *     } else {
 *         $off2 = mgsort2($offset, $runs-($runs/2), $base, $list2, $list1)
 *         mgsort2($off2, $runs/2, $base, $list2, $list1)
 *         merge the adjacent runs at $offset of $list1 into $list2
 *         return the offset of the end of the merged runs
 *     }
 * }
 * mgsort2(0, $runs, $base, $aux, $base);
 *
 * For our 5 runs, the tree of calls looks like 
 *
 *           5
 *      3        2
 *   2     1   1   1
 * 1   1
 *
 * 1   2   3   4   5
 *
 * and the corresponding activity looks like
 *
 * copy runs 1 and 2 from base to aux
 * merge runs 1 and 2 from aux to base
 * (run 3 is where it belongs, no copy needed)
 * merge runs 12 and 3 from base to aux
 * (runs 4 and 5 are where they belong, no copy needed)
 * merge runs 4 and 5 from base to aux
 * merge runs 123 and 45 from aux to base
 *
 * Note that we merge runs 1 and 2 immediately after copying them,
 * while they are still likely to be in fast cache.  Similarly,
 * run 3 is merged with run 12 while it still may be lingering in cache.
 * This implementation should therefore enjoy much of the cache-friendly
 * behavior that quicksort does.  In addition, it does less copying
 * than the original mergesort implementation (only runs 1 and 2 are copied)
 * and the "balancing" of merges is better (merged runs comprise more nearly
 * equal numbers of original runs).
 *
 * The actual cache-friendly implementation will use a pseudo-stack
 * to avoid recursion, and will unroll processing of runs of length 2,
 * but it is otherwise similar to the recursive implementation.
 */

typedef struct {
    IV	offset;		/* offset of 1st of 2 runs at this level */
    IV	runs;		/* how many runs must be combined into 1 */
} off_runs;		/* pseudo-stack element */


static I32
cmp_desc(pTHX_ gptr const a, gptr const b)
{
    dVAR;
    return -PL_sort_RealCmp(aTHX_ a, b);
}

STATIC void
S_mergesortsv(pTHX_ gptr *base, size_t nmemb, SVCOMPARE_t cmp, U32 flags)
{
    dVAR;
    IV i, run, offset;
    I32 sense, level;
    register gptr *f1, *f2, *t, *b, *p;
    int iwhich;
    gptr *aux;
    gptr *p1;
    gptr small[SMALLSORT];
    gptr *which[3];
    off_runs stack[60], *stackp;
    SVCOMPARE_t savecmp = NULL;

    if (nmemb <= 1) return;			/* sorted trivially */

    if ((flags & SORTf_DESC) != 0) {
	savecmp = PL_sort_RealCmp;	/* Save current comparison routine, if any */
	PL_sort_RealCmp = cmp;	/* Put comparison routine where cmp_desc can find it */
	cmp = cmp_desc;
    }

    if (nmemb <= SMALLSORT) aux = small;	/* use stack for aux array */
    else { Newx(aux,nmemb,gptr); }		/* allocate auxilliary array */
    level = 0;
    stackp = stack;
    stackp->runs = dynprep(aTHX_ base, aux, nmemb, cmp);
    stackp->offset = offset = 0;
    which[0] = which[2] = base;
    which[1] = aux;
    for (;;) {
	/* On levels where both runs have be constructed (stackp->runs == 0),
	 * merge them, and note the offset of their end, in case the offset
	 * is needed at the next level up.  Hop up a level, and,
	 * as long as stackp->runs is 0, keep merging.
	 */
	IV runs = stackp->runs;
	if (runs == 0) {
	    gptr *list1, *list2;
	    iwhich = level & 1;
	    list1 = which[iwhich];		/* area where runs are now */
	    list2 = which[++iwhich];		/* area for merged runs */
	    do {
		register gptr *l1, *l2, *tp2;
		offset = stackp->offset;
		f1 = p1 = list1 + offset;		/* start of first run */
		p = tp2 = list2 + offset;	/* where merged run will go */
		t = NEXT(p);			/* where first run ends */
		f2 = l1 = POTHER(t, list2, list1); /* ... on the other side */
		t = NEXT(t);			/* where second runs ends */
		l2 = POTHER(t, list2, list1);	/* ... on the other side */
		offset = PNELEM(list2, t);
		while (f1 < l1 && f2 < l2) {
		    /* If head 1 is larger than head 2, find ALL the elements
		    ** in list 2 strictly less than head1, write them all,
		    ** then head 1.  Then compare the new heads, and repeat,
		    ** until one or both lists are exhausted.
		    **
		    ** In all comparisons (after establishing
		    ** which head to merge) the item to merge
		    ** (at pointer q) is the first operand of
		    ** the comparison.  When we want to know
		    ** if "q is strictly less than the other",
		    ** we can't just do
		    **    cmp(q, other) < 0
		    ** because stability demands that we treat equality
		    ** as high when q comes from l2, and as low when
		    ** q was from l1.  So we ask the question by doing
		    **    cmp(q, other) <= sense
		    ** and make sense == 0 when equality should look low,
		    ** and -1 when equality should look high.
		    */

		    register gptr *q;
		    if (cmp(aTHX_ *f1, *f2) <= 0) {
			q = f2; b = f1; t = l1;
			sense = -1;
		    } else {
			q = f1; b = f2; t = l2;
			sense = 0;
		    }


		    /* ramp up
		    **
		    ** Leave t at something strictly
		    ** greater than q (or at the end of the list),
		    ** and b at something strictly less than q.
		    */
		    for (i = 1, run = 0 ;;) {
			if ((p = PINDEX(b, i)) >= t) {
			    /* off the end */
			    if (((p = PINDEX(t, -1)) > b) &&
				(cmp(aTHX_ *q, *p) <= sense))
				 t = p;
			    else b = p;
			    break;
			} else if (cmp(aTHX_ *q, *p) <= sense) {
			    t = p;
			    break;
			} else b = p;
			if (++run >= RTHRESH) i += i;
		    }


		    /* q is known to follow b and must be inserted before t.
		    ** Increment b, so the range of possibilities is [b,t).
		    ** Round binary split down, to favor early appearance.
		    ** Adjust b and t until q belongs just before t.
		    */

		    b++;
		    while (b < t) {
			p = PINDEX(b, (PNELEM(b, t) - 1) / 2);
			if (cmp(aTHX_ *q, *p) <= sense) {
			    t = p;
			} else b = p + 1;
		    }


		    /* Copy all the strictly low elements */

		    if (q == f1) {
			FROMTOUPTO(f2, tp2, t);
			*tp2++ = *f1++;
		    } else {
			FROMTOUPTO(f1, tp2, t);
			*tp2++ = *f2++;
		    }
		}


		/* Run out remaining list */
		if (f1 == l1) {
		       if (f2 < l2) FROMTOUPTO(f2, tp2, l2);
		} else              FROMTOUPTO(f1, tp2, l1);
		p1 = NEXT(p1) = POTHER(tp2, list2, list1);

		if (--level == 0) goto done;
		--stackp;
		t = list1; list1 = list2; list2 = t;	/* swap lists */
	    } while ((runs = stackp->runs) == 0);
	}


	stackp->runs = 0;		/* current run will finish level */
	/* While there are more than 2 runs remaining,
	 * turn them into exactly 2 runs (at the "other" level),
	 * each made up of approximately half the runs.
	 * Stack the second half for later processing,
	 * and set about producing the first half now.
	 */
	while (runs > 2) {
	    ++level;
	    ++stackp;
	    stackp->offset = offset;
	    runs -= stackp->runs = runs / 2;
	}
	/* We must construct a single run from 1 or 2 runs.
	 * All the original runs are in which[0] == base.
	 * The run we construct must end up in which[level&1].
	 */
	iwhich = level & 1;
	if (runs == 1) {
	    /* Constructing a single run from a single run.
	     * If it's where it belongs already, there's nothing to do.
	     * Otherwise, copy it to where it belongs.
	     * A run of 1 is either a singleton at level 0,
	     * or the second half of a split 3.  In neither event
	     * is it necessary to set offset.  It will be set by the merge
	     * that immediately follows.
	     */
	    if (iwhich) {	/* Belongs in aux, currently in base */
		f1 = b = PINDEX(base, offset);	/* where list starts */
		f2 = PINDEX(aux, offset);	/* where list goes */
		t = NEXT(f2);			/* where list will end */
		offset = PNELEM(aux, t);	/* offset thereof */
		t = PINDEX(base, offset);	/* where it currently ends */
		FROMTOUPTO(f1, f2, t);		/* copy */
		NEXT(b) = t;			/* set up parallel pointer */
	    } else if (level == 0) goto done;	/* single run at level 0 */
	} else {
	    /* Constructing a single run from two runs.
	     * The merge code at the top will do that.
	     * We need only make sure the two runs are in the "other" array,
	     * so they'll end up in the correct array after the merge.
	     */
	    ++level;
	    ++stackp;
	    stackp->offset = offset;
	    stackp->runs = 0;	/* take care of both runs, trigger merge */
	    if (!iwhich) {	/* Merged runs belong in aux, copy 1st */
		f1 = b = PINDEX(base, offset);	/* where first run starts */
		f2 = PINDEX(aux, offset);	/* where it will be copied */
		t = NEXT(f2);			/* where first run will end */
		offset = PNELEM(aux, t);	/* offset thereof */
		p = PINDEX(base, offset);	/* end of first run */
		t = NEXT(t);			/* where second run will end */
		t = PINDEX(base, PNELEM(aux, t)); /* where it now ends */
		FROMTOUPTO(f1, f2, t);		/* copy both runs */
		NEXT(b) = p;			/* paralled pointer for 1st */
		NEXT(p) = t;			/* ... and for second */
	    }
	}
    }
done:
    if (aux != small) Safefree(aux);	/* free iff allocated */
    if (flags) {
	 PL_sort_RealCmp = savecmp;	/* Restore current comparison routine, if any */
    }
    return;
}

/*
 * The quicksort implementation was derived from source code contributed
 * by Tom Horsley.
 *
 * NOTE: this code was derived from Tom Horsley's qsort replacement
 * and should not be confused with the original code.
 */

/* Copyright (C) Tom Horsley, 1997. All rights reserved.

   Permission granted to distribute under the same terms as perl which are
   (briefly):

    This program is free software; you can redistribute it and/or modify
    it under the terms of either:

	a) the GNU General Public License as published by the Free
	Software Foundation; either version 1, or (at your option) any
	later version, or

	b) the "Artistic License" which comes with this Kit.

   Details on the perl license can be found in the perl source code which
   may be located via the www.perl.com web page.

   This is the most wonderfulest possible qsort I can come up with (and
   still be mostly portable) My (limited) tests indicate it consistently
   does about 20% fewer calls to compare than does the qsort in the Visual
   C++ library, other vendors may vary.

   Some of the ideas in here can be found in "Algorithms" by Sedgewick,
   others I invented myself (or more likely re-invented since they seemed
   pretty obvious once I watched the algorithm operate for a while).

   Most of this code was written while watching the Marlins sweep the Giants
   in the 1997 National League Playoffs - no Braves fans allowed to use this
   code (just kidding :-).

   I realize that if I wanted to be true to the perl tradition, the only
   comment in this file would be something like:

   ...they shuffled back towards the rear of the line. 'No, not at the
   rear!'  the slave-driver shouted. 'Three files up. And stay there...

   However, I really needed to violate that tradition just so I could keep
   track of what happens myself, not to mention some poor fool trying to
   understand this years from now :-).
*/

/* ********************************************************** Configuration */

#ifndef QSORT_ORDER_GUESS
#define QSORT_ORDER_GUESS 2	/* Select doubling version of the netBSD trick */
#endif

/* QSORT_MAX_STACK is the largest number of partitions that can be stacked up for
   future processing - a good max upper bound is log base 2 of memory size
   (32 on 32 bit machines, 64 on 64 bit machines, etc). In reality can
   safely be smaller than that since the program is taking up some space and
   most operating systems only let you grab some subset of contiguous
   memory (not to mention that you are normally sorting data larger than
   1 byte element size :-).
*/
#ifndef QSORT_MAX_STACK
#define QSORT_MAX_STACK 32
#endif

/* QSORT_BREAK_EVEN is the size of the largest partition we should insertion sort.
   Anything bigger and we use qsort. If you make this too small, the qsort
   will probably break (or become less efficient), because it doesn't expect
   the middle element of a partition to be the same as the right or left -
   you have been warned).
*/
#ifndef QSORT_BREAK_EVEN
#define QSORT_BREAK_EVEN 6
#endif

/* QSORT_PLAY_SAFE is the size of the largest partition we're willing
   to go quadratic on.  We innoculate larger partitions against
   quadratic behavior by shuffling them before sorting.  This is not
   an absolute guarantee of non-quadratic behavior, but it would take
   staggeringly bad luck to pick extreme elements as the pivot
   from randomized data.
*/
#ifndef QSORT_PLAY_SAFE
#define QSORT_PLAY_SAFE 255
#endif

/* ************************************************************* Data Types */

/* hold left and right index values of a partition waiting to be sorted (the
   partition includes both left and right - right is NOT one past the end or
   anything like that).
*/
struct partition_stack_entry {
   int left;
   int right;
#ifdef QSORT_ORDER_GUESS
   int qsort_break_even;
#endif
};

/* ******************************************************* Shorthand Macros */

/* Note that these macros will be used from inside the qsort function where
   we happen to know that the variable 'elt_size' contains the size of an
   array element and the variable 'temp' points to enough space to hold a
   temp element and the variable 'array' points to the array being sorted
   and 'compare' is the pointer to the compare routine.

   Also note that there are very many highly architecture specific ways
   these might be sped up, but this is simply the most generally portable
   code I could think of.
*/

/* Return < 0 == 0 or > 0 as the value of elt1 is < elt2, == elt2, > elt2
*/
#define qsort_cmp(elt1, elt2) \
   ((*compare)(aTHX_ array[elt1], array[elt2]))

#ifdef QSORT_ORDER_GUESS
#define QSORT_NOTICE_SWAP swapped++;
#else
#define QSORT_NOTICE_SWAP
#endif

/* swaps contents of array elements elt1, elt2.
*/
#define qsort_swap(elt1, elt2) \
   STMT_START { \
      QSORT_NOTICE_SWAP \
      temp = array[elt1]; \
      array[elt1] = array[elt2]; \
      array[elt2] = temp; \
   } STMT_END

/* rotate contents of elt1, elt2, elt3 such that elt1 gets elt2, elt2 gets
   elt3 and elt3 gets elt1.
*/
#define qsort_rotate(elt1, elt2, elt3) \
   STMT_START { \
      QSORT_NOTICE_SWAP \
      temp = array[elt1]; \
      array[elt1] = array[elt2]; \
      array[elt2] = array[elt3]; \
      array[elt3] = temp; \
   } STMT_END

/* ************************************************************ Debug stuff */

#ifdef QSORT_DEBUG

static void
break_here()
{
   return; /* good place to set a breakpoint */
}

#define qsort_assert(t) (void)( (t) || (break_here(), 0) )

static void
doqsort_all_asserts(
   void * array,
   size_t num_elts,
   size_t elt_size,
   int (*compare)(const void * elt1, const void * elt2),
   int pc_left, int pc_right, int u_left, int u_right)
{
   int i;

   qsort_assert(pc_left <= pc_right);
   qsort_assert(u_right < pc_left);
   qsort_assert(pc_right < u_left);
   for (i = u_right + 1; i < pc_left; ++i) {
      qsort_assert(qsort_cmp(i, pc_left) < 0);
   }
   for (i = pc_left; i < pc_right; ++i) {
      qsort_assert(qsort_cmp(i, pc_right) == 0);
   }
   for (i = pc_right + 1; i < u_left; ++i) {
      qsort_assert(qsort_cmp(pc_right, i) < 0);
   }
}

#define qsort_all_asserts(PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT) \
   doqsort_all_asserts(array, num_elts, elt_size, compare, \
                 PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT)

#else

#define qsort_assert(t) ((void)0)

#define qsort_all_asserts(PC_LEFT, PC_RIGHT, U_LEFT, U_RIGHT) ((void)0)

#endif

/* ****************************************************************** qsort */

STATIC void /* the standard unstable (u) quicksort (qsort) */
S_qsortsvu(pTHX_ SV ** array, size_t num_elts, SVCOMPARE_t compare)
{
   register SV * temp;
   struct partition_stack_entry partition_stack[QSORT_MAX_STACK];
   int next_stack_entry = 0;
   int part_left;
   int part_right;
#ifdef QSORT_ORDER_GUESS
   int qsort_break_even;
   int swapped;
#endif

    PERL_ARGS_ASSERT_QSORTSVU;

   /* Make sure we actually have work to do.
   */
   if (num_elts <= 1) {
      return;
   }

   /* Innoculate large partitions against quadratic behavior */
   if (num_elts > QSORT_PLAY_SAFE) {
      register size_t n;
      register SV ** const q = array;
      for (n = num_elts; n > 1; ) {
         register const size_t j = (size_t)(n-- * Drand01());
         temp = q[j];
         q[j] = q[n];
         q[n] = temp;
      }
   }

   /* Setup the initial partition definition and fall into the sorting loop
   */
   part_left = 0;
   part_right = (int)(num_elts - 1);
#ifdef QSORT_ORDER_GUESS
   qsort_break_even = QSORT_BREAK_EVEN;
#else
#define qsort_break_even QSORT_BREAK_EVEN
#endif
   for ( ; ; ) {
      if ((part_right - part_left) >= qsort_break_even) {
         /* OK, this is gonna get hairy, so lets try to document all the
            concepts and abbreviations and variables and what they keep
            track of:

            pc: pivot chunk - the set of array elements we accumulate in the
                middle of the partition, all equal in value to the original
                pivot element selected. The pc is defined by:

                pc_left - the leftmost array index of the pc
                pc_right - the rightmost array index of the pc

                we start with pc_left == pc_right and only one element
                in the pivot chunk (but it can grow during the scan).

            u:  uncompared elements - the set of elements in the partition
                we have not yet compared to the pivot value. There are two
                uncompared sets during the scan - one to the left of the pc
                and one to the right.

                u_right - the rightmost index of the left side's uncompared set
                u_left - the leftmost index of the right side's uncompared set

                The leftmost index of the left sides's uncompared set
                doesn't need its own variable because it is always defined
                by the leftmost edge of the whole partition (part_left). The
                same goes for the rightmost edge of the right partition
                (part_right).

                We know there are no uncompared elements on the left once we
                get u_right < part_left and no uncompared elements on the
                right once u_left > part_right. When both these conditions
                are met, we have completed the scan of the partition.

                Any elements which are between the pivot chunk and the
                uncompared elements should be less than the pivot value on
                the left side and greater than the pivot value on the right
                side (in fact, the goal of the whole algorithm is to arrange
                for that to be true and make the groups of less-than and
                greater-then elements into new partitions to sort again).

            As you marvel at the complexity of the code and wonder why it
            has to be so confusing. Consider some of the things this level
            of confusion brings:

            Once I do a compare, I squeeze every ounce of juice out of it. I
            never do compare calls I don't have to do, and I certainly never
            do redundant calls.

            I also never swap any elements unless I can prove there is a
            good reason. Many sort algorithms will swap a known value with
            an uncompared value just to get things in the right place (or
            avoid complexity :-), but that uncompared value, once it gets
            compared, may then have to be swapped again. A lot of the
            complexity of this code is due to the fact that it never swaps
            anything except compared values, and it only swaps them when the
            compare shows they are out of position.
         */
         int pc_left, pc_right;
         int u_right, u_left;

         int s;

         pc_left = ((part_left + part_right) / 2);
         pc_right = pc_left;
         u_right = pc_left - 1;
         u_left = pc_right + 1;

         /* Qsort works best when the pivot value is also the median value
            in the partition (unfortunately you can't find the median value
            without first sorting :-), so to give the algorithm a helping
            hand, we pick 3 elements and sort them and use the median value
            of that tiny set as the pivot value.

            Some versions of qsort like to use the left middle and right as
            the 3 elements to sort so they can insure the ends of the
            partition will contain values which will stop the scan in the
            compare loop, but when you have to call an arbitrarily complex
            routine to do a compare, its really better to just keep track of
            array index values to know when you hit the edge of the
            partition and avoid the extra compare. An even better reason to
            avoid using a compare call is the fact that you can drop off the
            edge of the array if someone foolishly provides you with an
            unstable compare function that doesn't always provide consistent
            results.

            So, since it is simpler for us to compare the three adjacent
            elements in the middle of the partition, those are the ones we
            pick here (conveniently pointed at by u_right, pc_left, and
            u_left). The values of the left, center, and right elements
            are refered to as l c and r in the following comments.
         */

#ifdef QSORT_ORDER_GUESS
         swapped = 0;
#endif
         s = qsort_cmp(u_right, pc_left);
         if (s < 0) {
            /* l < c */
            s = qsort_cmp(pc_left, u_left);
            /* if l < c, c < r - already in order - nothing to do */
            if (s == 0) {
               /* l < c, c == r - already in order, pc grows */
               ++pc_right;
               qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
            } else if (s > 0) {
               /* l < c, c > r - need to know more */
               s = qsort_cmp(u_right, u_left);
               if (s < 0) {
                  /* l < c, c > r, l < r - swap c & r to get ordered */
                  qsort_swap(pc_left, u_left);
                  qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
               } else if (s == 0) {
                  /* l < c, c > r, l == r - swap c&r, grow pc */
                  qsort_swap(pc_left, u_left);
                  --pc_left;
                  qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
               } else {
                  /* l < c, c > r, l > r - make lcr into rlc to get ordered */
                  qsort_rotate(pc_left, u_right, u_left);
                  qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
               }
            }
         } else if (s == 0) {
            /* l == c */
            s = qsort_cmp(pc_left, u_left);
            if (s < 0) {
               /* l == c, c < r - already in order, grow pc */
               --pc_left;
               qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
            } else if (s == 0) {
               /* l == c, c == r - already in order, grow pc both ways */
               --pc_left;
               ++pc_right;
               qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
            } else {
               /* l == c, c > r - swap l & r, grow pc */
               qsort_swap(u_right, u_left);
               ++pc_right;
               qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
            }
         } else {
            /* l > c */
            s = qsort_cmp(pc_left, u_left);
            if (s < 0) {
               /* l > c, c < r - need to know more */
               s = qsort_cmp(u_right, u_left);
               if (s < 0) {
                  /* l > c, c < r, l < r - swap l & c to get ordered */
                  qsort_swap(u_right, pc_left);
                  qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
               } else if (s == 0) {
                  /* l > c, c < r, l == r - swap l & c, grow pc */
                  qsort_swap(u_right, pc_left);
                  ++pc_right;
                  qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
               } else {
                  /* l > c, c < r, l > r - rotate lcr into crl to order */
                  qsort_rotate(u_right, pc_left, u_left);
                  qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
               }
            } else if (s == 0) {
               /* l > c, c == r - swap ends, grow pc */
               qsort_swap(u_right, u_left);
               --pc_left;
               qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
            } else {
               /* l > c, c > r - swap ends to get in order */
               qsort_swap(u_right, u_left);
               qsort_all_asserts(pc_left, pc_right, u_left + 1, u_right - 1);
            }
         }
         /* We now know the 3 middle elements have been compared and
            arranged in the desired order, so we can shrink the uncompared
            sets on both sides
         */
         --u_right;
         ++u_left;
         qsort_all_asserts(pc_left, pc_right, u_left, u_right);

         /* The above massive nested if was the simple part :-). We now have
            the middle 3 elements ordered and we need to scan through the
            uncompared sets on either side, swapping elements that are on
            the wrong side or simply shuffling equal elements around to get
            all equal elements into the pivot chunk.
         */

         for ( ; ; ) {
            int still_work_on_left;
            int still_work_on_right;

            /* Scan the uncompared values on the left. If I find a value
               equal to the pivot value, move it over so it is adjacent to
               the pivot chunk and expand the pivot chunk. If I find a value
               less than the pivot value, then just leave it - its already
               on the correct side of the partition. If I find a greater
               value, then stop the scan.
            */
            while ((still_work_on_left = (u_right >= part_left))) {
               s = qsort_cmp(u_right, pc_left);
               if (s < 0) {
                  --u_right;
               } else if (s == 0) {
                  --pc_left;
                  if (pc_left != u_right) {
                     qsort_swap(u_right, pc_left);
                  }
                  --u_right;
               } else {
                  break;
               }
               qsort_assert(u_right < pc_left);
               qsort_assert(pc_left <= pc_right);
               qsort_assert(qsort_cmp(u_right + 1, pc_left) <= 0);
               qsort_assert(qsort_cmp(pc_left, pc_right) == 0);
            }

            /* Do a mirror image scan of uncompared values on the right
            */
            while ((still_work_on_right = (u_left <= part_right))) {
               s = qsort_cmp(pc_right, u_left);
               if (s < 0) {
                  ++u_left;
               } else if (s == 0) {
                  ++pc_right;
                  if (pc_right != u_left) {
                     qsort_swap(pc_right, u_left);
                  }
                  ++u_left;
               } else {
                  break;
               }
               qsort_assert(u_left > pc_right);
               qsort_assert(pc_left <= pc_right);
               qsort_assert(qsort_cmp(pc_right, u_left - 1) <= 0);
               qsort_assert(qsort_cmp(pc_left, pc_right) == 0);
            }

            if (still_work_on_left) {
               /* I know I have a value on the left side which needs to be
                  on the right side, but I need to know more to decide
                  exactly the best thing to do with it.
               */
               if (still_work_on_right) {
                  /* I know I have values on both side which are out of
                     position. This is a big win because I kill two birds
                     with one swap (so to speak). I can advance the
                     uncompared pointers on both sides after swapping both
                     of them into the right place.
                  */
                  qsort_swap(u_right, u_left);
                  --u_right;
                  ++u_left;
                  qsort_all_asserts(pc_left, pc_right, u_left, u_right);
               } else {
                  /* I have an out of position value on the left, but the
                     right is fully scanned, so I "slide" the pivot chunk
                     and any less-than values left one to make room for the
                     greater value over on the right. If the out of position
                     value is immediately adjacent to the pivot chunk (there
                     are no less-than values), I can do that with a swap,
                     otherwise, I have to rotate one of the less than values
                     into the former position of the out of position value
                     and the right end of the pivot chunk into the left end
                     (got all that?).
                  */
                  --pc_left;
                  if (pc_left == u_right) {
                     qsort_swap(u_right, pc_right);
                     qsort_all_asserts(pc_left, pc_right-1, u_left, u_right-1);
                  } else {
                     qsort_rotate(u_right, pc_left, pc_right);
                     qsort_all_asserts(pc_left, pc_right-1, u_left, u_right-1);
                  }
                  --pc_right;
                  --u_right;
               }
            } else if (still_work_on_right) {
               /* Mirror image of complex case above: I have an out of
                  position value on the right, but the left is fully
                  scanned, so I need to shuffle things around to make room
                  for the right value on the left.
               */
               ++pc_right;
               if (pc_right == u_left) {
                  qsort_swap(u_left, pc_left);
                  qsort_all_asserts(pc_left+1, pc_right, u_left+1, u_right);
               } else {
                  qsort_rotate(pc_right, pc_left, u_left);
                  qsort_all_asserts(pc_left+1, pc_right, u_left+1, u_right);
               }
               ++pc_left;
               ++u_left;
            } else {
               /* No more scanning required on either side of partition,
                  break out of loop and figure out next set of partitions
               */
               break;
            }
         }

         /* The elements in the pivot chunk are now in the right place. They
            will never move or be compared again. All I have to do is decide
            what to do with the stuff to the left and right of the pivot
            chunk.

            Notes on the QSORT_ORDER_GUESS ifdef code:

            1. If I just built these partitions without swapping any (or
               very many) elements, there is a chance that the elements are
               already ordered properly (being properly ordered will
               certainly result in no swapping, but the converse can't be
               proved :-).

            2. A (properly written) insertion sort will run faster on
               already ordered data than qsort will.

            3. Perhaps there is some way to make a good guess about
               switching to an insertion sort earlier than partition size 6
               (for instance - we could save the partition size on the stack
               and increase the size each time we find we didn't swap, thus
               switching to insertion sort earlier for partitions with a
               history of not swapping).

            4. Naturally, if I just switch right away, it will make
               artificial benchmarks with pure ascending (or descending)
               data look really good, but is that a good reason in general?
               Hard to say...
         */

#ifdef QSORT_ORDER_GUESS
         if (swapped < 3) {
#if QSORT_ORDER_GUESS == 1
            qsort_break_even = (part_right - part_left) + 1;
#endif
#if QSORT_ORDER_GUESS == 2
            qsort_break_even *= 2;
#endif
#if QSORT_ORDER_GUESS == 3
            const int prev_break = qsort_break_even;
            qsort_break_even *= qsort_break_even;
            if (qsort_break_even < prev_break) {
               qsort_break_even = (part_right - part_left) + 1;
            }
#endif
         } else {
            qsort_break_even = QSORT_BREAK_EVEN;
         }
#endif

         if (part_left < pc_left) {
            /* There are elements on the left which need more processing.
               Check the right as well before deciding what to do.
            */
            if (pc_right < part_right) {
               /* We have two partitions to be sorted. Stack the biggest one
                  and process the smallest one on the next iteration. This
                  minimizes the stack height by insuring that any additional
                  stack entries must come from the smallest partition which
                  (because it is smallest) will have the fewest
                  opportunities to generate additional stack entries.
               */
               if ((part_right - pc_right) > (pc_left - part_left)) {
                  /* stack the right partition, process the left */
                  partition_stack[next_stack_entry].left = pc_right + 1;
                  partition_stack[next_stack_entry].right = part_right;
#ifdef QSORT_ORDER_GUESS
                  partition_stack[next_stack_entry].qsort_break_even = qsort_break_even;
#endif
                  part_right = pc_left - 1;
               } else {
                  /* stack the left partition, process the right */
                  partition_stack[next_stack_entry].left = part_left;
                  partition_stack[next_stack_entry].right = pc_left - 1;
#ifdef QSORT_ORDER_GUESS
                  partition_stack[next_stack_entry].qsort_break_even = qsort_break_even;
#endif
                  part_left = pc_right + 1;
               }
               qsort_assert(next_stack_entry < QSORT_MAX_STACK);
               ++next_stack_entry;
            } else {
               /* The elements on the left are the only remaining elements
                  that need sorting, arrange for them to be processed as the
                  next partition.
               */
               part_right = pc_left - 1;
            }
         } else if (pc_right < part_right) {
            /* There is only one chunk on the right to be sorted, make it
               the new partition and loop back around.
            */
            part_left = pc_right + 1;
         } else {
            /* This whole partition wound up in the pivot chunk, so
               we need to get a new partition off the stack.
            */
            if (next_stack_entry == 0) {
               /* the stack is empty - we are done */
               break;
            }
            --next_stack_entry;
            part_left = partition_stack[next_stack_entry].left;
            part_right = partition_stack[next_stack_entry].right;
#ifdef QSORT_ORDER_GUESS
            qsort_break_even = partition_stack[next_stack_entry].qsort_break_even;
#endif
         }
      } else {
         /* This partition is too small to fool with qsort complexity, just
            do an ordinary insertion sort to minimize overhead.
         */
         int i;
         /* Assume 1st element is in right place already, and start checking
            at 2nd element to see where it should be inserted.
         */
         for (i = part_left + 1; i <= part_right; ++i) {
            int j;
            /* Scan (backwards - just in case 'i' is already in right place)
               through the elements already sorted to see if the ith element
               belongs ahead of one of them.
            */
            for (j = i - 1; j >= part_left; --j) {
               if (qsort_cmp(i, j) >= 0) {
                  /* i belongs right after j
                  */
                  break;
               }
            }
            ++j;
            if (j != i) {
               /* Looks like we really need to move some things
               */
	       int k;
	       temp = array[i];
	       for (k = i - 1; k >= j; --k)
		  array[k + 1] = array[k];
               array[j] = temp;
            }
         }

         /* That partition is now sorted, grab the next one, or get out
            of the loop if there aren't any more.
         */

         if (next_stack_entry == 0) {
            /* the stack is empty - we are done */
            break;
         }
         --next_stack_entry;
         part_left = partition_stack[next_stack_entry].left;
         part_right = partition_stack[next_stack_entry].right;
#ifdef QSORT_ORDER_GUESS
         qsort_break_even = partition_stack[next_stack_entry].qsort_break_even;
#endif
      }
   }

   /* Believe it or not, the array is sorted at this point! */
}

/* Stabilize what is, presumably, an otherwise unstable sort method.
 * We do that by allocating (or having on hand) an array of pointers
 * that is the same size as the original array of elements to be sorted.
 * We initialize this parallel array with the addresses of the original
 * array elements.  This indirection can make you crazy.
 * Some pictures can help.  After initializing, we have
 *
 *  indir                  list1
 * +----+                 +----+
 * |    | --------------> |    | ------> first element to be sorted
 * +----+                 +----+
 * |    | --------------> |    | ------> second element to be sorted
 * +----+                 +----+
 * |    | --------------> |    | ------> third element to be sorted
 * +----+                 +----+
 *  ...
 * +----+                 +----+
 * |    | --------------> |    | ------> n-1st element to be sorted
 * +----+                 +----+
 * |    | --------------> |    | ------> n-th element to be sorted
 * +----+                 +----+
 *
 * During the sort phase, we leave the elements of list1 where they are,
 * and sort the pointers in the indirect array in the same order determined
 * by the original comparison routine on the elements pointed to.
 * Because we don't move the elements of list1 around through
 * this phase, we can break ties on elements that compare equal
 * using their address in the list1 array, ensuring stability.
 * This leaves us with something looking like
 *
 *  indir                  list1
 * +----+                 +----+
 * |    | --+       +---> |    | ------> first element to be sorted
 * +----+   |       |     +----+
 * |    | --|-------|---> |    | ------> second element to be sorted
 * +----+   |       |     +----+
 * |    | --|-------+ +-> |    | ------> third element to be sorted
 * +----+   |         |   +----+
 *  ...
 * +----+    | |   | |    +----+
 * |    | ---|-+   | +--> |    | ------> n-1st element to be sorted
 * +----+    |     |      +----+
 * |    | ---+     +----> |    | ------> n-th element to be sorted
 * +----+                 +----+
 *
 * where the i-th element of the indirect array points to the element
 * that should be i-th in the sorted array.  After the sort phase,
 * we have to put the elements of list1 into the places
 * dictated by the indirect array.
 */


static I32
cmpindir(pTHX_ gptr const a, gptr const b)
{
    dVAR;
    gptr * const ap = (gptr *)a;
    gptr * const bp = (gptr *)b;
    const I32 sense = PL_sort_RealCmp(aTHX_ *ap, *bp);

    if (sense)
	return sense;
    return (ap > bp) ? 1 : ((ap < bp) ? -1 : 0);
}

static I32
cmpindir_desc(pTHX_ gptr const a, gptr const b)
{
    dVAR;
    gptr * const ap = (gptr *)a;
    gptr * const bp = (gptr *)b;
    const I32 sense = PL_sort_RealCmp(aTHX_ *ap, *bp);

    /* Reverse the default */
    if (sense)
	return -sense;
    /* But don't reverse the stability test.  */
    return (ap > bp) ? 1 : ((ap < bp) ? -1 : 0);

}

STATIC void
S_qsortsv(pTHX_ gptr *list1, size_t nmemb, SVCOMPARE_t cmp, U32 flags)
{
    dVAR;
    if ((flags & SORTf_STABLE) != 0) {
	 register gptr **pp, *q;
	 register size_t n, j, i;
	 gptr *small[SMALLSORT], **indir, tmp;
	 SVCOMPARE_t savecmp;
	 if (nmemb <= 1) return;     /* sorted trivially */

	 /* Small arrays can use the stack, big ones must be allocated */
	 if (nmemb <= SMALLSORT) indir = small;
	 else { Newx(indir, nmemb, gptr *); }

	 /* Copy pointers to original array elements into indirect array */
	 for (n = nmemb, pp = indir, q = list1; n--; ) *pp++ = q++;

	 savecmp = PL_sort_RealCmp;	/* Save current comparison routine, if any */
	 PL_sort_RealCmp = cmp;	/* Put comparison routine where cmpindir can find it */

	 /* sort, with indirection */
	 if (flags & SORTf_DESC)
	    qsortsvu((gptr *)indir, nmemb, cmpindir_desc);
	else
	    qsortsvu((gptr *)indir, nmemb, cmpindir);

	 pp = indir;
	 q = list1;
	 for (n = nmemb; n--; ) {
	      /* Assert A: all elements of q with index > n are already
	       * in place.  This is vacuosly true at the start, and we
	       * put element n where it belongs below (if it wasn't
	       * already where it belonged). Assert B: we only move
	       * elements that aren't where they belong,
	       * so, by A, we never tamper with elements above n.
	       */
	      j = pp[n] - q;		/* This sets j so that q[j] is
					 * at pp[n].  *pp[j] belongs in
					 * q[j], by construction.
					 */
	      if (n != j) {		/* all's well if n == j */
		   tmp = q[j];		/* save what's in q[j] */
		   do {
			q[j] = *pp[j];	/* put *pp[j] where it belongs */
			i = pp[j] - q;	/* the index in q of the element
					 * just moved */
			pp[j] = q + j;	/* this is ok now */
		   } while ((j = i) != n);
		   /* There are only finitely many (nmemb) addresses
		    * in the pp array.
		    * So we must eventually revisit an index we saw before.
		    * Suppose the first revisited index is k != n.
		    * An index is visited because something else belongs there.
		    * If we visit k twice, then two different elements must
		    * belong in the same place, which cannot be.
		    * So j must get back to n, the loop terminates,
		    * and we put the saved element where it belongs.
		    */
		   q[n] = tmp;		/* put what belongs into
					 * the n-th element */
	      }
	 }

	/* free iff allocated */
	 if (indir != small) { Safefree(indir); }
	 /* restore prevailing comparison routine */
	 PL_sort_RealCmp = savecmp;
    } else if ((flags & SORTf_DESC) != 0) {
	 const SVCOMPARE_t savecmp = PL_sort_RealCmp;	/* Save current comparison routine, if any */
	 PL_sort_RealCmp = cmp;	/* Put comparison routine where cmp_desc can find it */
	 cmp = cmp_desc;
	 qsortsvu(list1, nmemb, cmp);
	 /* restore prevailing comparison routine */
	 PL_sort_RealCmp = savecmp;
    } else {
	 qsortsvu(list1, nmemb, cmp);
    }
}

/*
=head1 Array Manipulation Functions

=for apidoc sortsv

Sort an array. Here is an example:

    sortsv(AvARRAY(av), av_len(av)+1, Perl_sv_cmp_locale);

Currently this always uses mergesort. See sortsv_flags for a more
flexible routine.

=cut
*/

void
Perl_sortsv(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp)
{
    PERL_ARGS_ASSERT_SORTSV;

    sortsv_flags(array, nmemb, cmp, 0);
}

/*
=for apidoc sortsv_flags

Sort an array, with various options.

=cut
*/
void
Perl_sortsv_flags(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp, U32 flags)
{
    PERL_ARGS_ASSERT_SORTSV_FLAGS;

    if (flags & SORTf_QSORT)
	S_qsortsv(aTHX_ array, nmemb, cmp, flags);
    else
	S_mergesortsv(aTHX_ array, nmemb, cmp, flags);
}

#define SvNSIOK(sv) ((SvFLAGS(sv) & SVf_NOK) || ((SvFLAGS(sv) & (SVf_IOK|SVf_IVisUV)) == SVf_IOK))
#define SvSIOK(sv) ((SvFLAGS(sv) & (SVf_IOK|SVf_IVisUV)) == SVf_IOK)
#define SvNSIV(sv) ( SvNOK(sv) ? SvNVX(sv) : ( SvSIOK(sv) ? SvIVX(sv) : sv_2nv(sv) ) )

PP(pp_sort)
{
    dVAR; dSP; dMARK; dORIGMARK;
    register SV **p1 = ORIGMARK+1, **p2;
    register I32 max, i;
    AV* av = NULL;
    HV *stash;
    GV *gv;
    CV *cv = NULL;
    I32 gimme = GIMME;
    OP* const nextop = PL_op->op_next;
    I32 overloading = 0;
    bool hasargs = FALSE;
    I32 is_xsub = 0;
    I32 sorting_av = 0;
    const U8 priv = PL_op->op_private;
    const U8 flags = PL_op->op_flags;
    U32 sort_flags = 0;
    void (*sortsvp)(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp, U32 flags)
      = Perl_sortsv_flags;
    I32 all_SIVs = 1;

    if ((priv & OPpSORT_DESCEND) != 0)
	sort_flags |= SORTf_DESC;
    if ((priv & OPpSORT_QSORT) != 0)
	sort_flags |= SORTf_QSORT;
    if ((priv & OPpSORT_STABLE) != 0)
	sort_flags |= SORTf_STABLE;

    if (gimme != G_ARRAY) {
	SP = MARK;
	EXTEND(SP,1);
	RETPUSHUNDEF;
    }

    ENTER;
    SAVEVPTR(PL_sortcop);
    if (flags & OPf_STACKED) {
	if (flags & OPf_SPECIAL) {
	    OP *kid = cLISTOP->op_first->op_sibling;	/* pass pushmark */
	    kid = kUNOP->op_first;			/* pass rv2gv */
	    kid = kUNOP->op_first;			/* pass leave */
	    PL_sortcop = kid->op_next;
	    stash = CopSTASH(PL_curcop);
	}
	else {
	    cv = sv_2cv(*++MARK, &stash, &gv, 0);
	    if (cv && SvPOK(cv)) {
		const char * const proto = SvPV_nolen_const(MUTABLE_SV(cv));
		if (proto && strEQ(proto, "$$")) {
		    hasargs = TRUE;
		}
	    }
	    if (!(cv && CvROOT(cv))) {
		if (cv && CvISXSUB(cv)) {
		    is_xsub = 1;
		}
		else if (gv) {
		    SV *tmpstr = sv_newmortal();
		    gv_efullname3(tmpstr, gv, NULL);
		    DIE(aTHX_ "Undefined sort subroutine \"%"SVf"\" called",
			SVfARG(tmpstr));
		}
		else {
		    DIE(aTHX_ "Undefined subroutine in sort");
		}
	    }

	    if (is_xsub)
		PL_sortcop = (OP*)cv;
	    else
		PL_sortcop = CvSTART(cv);
	}
    }
    else {
	PL_sortcop = NULL;
	stash = CopSTASH(PL_curcop);
    }

    /* optimiser converts "@a = sort @a" to "sort \@a";
     * in case of tied @a, pessimise: push (@a) onto stack, then assign
     * result back to @a at the end of this function */
    if (priv & OPpSORT_INPLACE) {
	assert( MARK+1 == SP && *SP && SvTYPE(*SP) == SVt_PVAV);
	(void)POPMARK; /* remove mark associated with ex-OP_AASSIGN */
	av = MUTABLE_AV((*SP));
	max = AvFILL(av) + 1;
	if (SvMAGICAL(av)) {
	    MEXTEND(SP, max);
	    for (i=0; i < max; i++) {
		SV **svp = av_fetch(av, i, FALSE);
		*SP++ = (svp) ? *svp : NULL;
	    }
	    SP--;
	    p1 = p2 = SP - (max-1);
	}
	else {
	    if (SvREADONLY(av))
		Perl_croak(aTHX_ "%s", PL_no_modify);
	    else
		SvREADONLY_on(av);
	    p1 = p2 = AvARRAY(av);
	    sorting_av = 1;
	}
    }
    else {
	p2 = MARK+1;
	max = SP - MARK;
   }

    /* shuffle stack down, removing optional initial cv (p1!=p2), plus
     * any nulls; also stringify or converting to integer or number as
     * required any args */
    for (i=max; i > 0 ; i--) {
	if ((*p1 = *p2++)) {			/* Weed out nulls. */
	    SvTEMP_off(*p1);
	    if (!PL_sortcop) {
		if (priv & OPpSORT_NUMERIC) {
		    if (priv & OPpSORT_INTEGER) {
			if (!SvIOK(*p1)) {
			    if (SvAMAGIC(*p1))
				overloading = 1;
			    else
				(void)sv_2iv(*p1);
			}
		    }
		    else {
			if (!SvNSIOK(*p1)) {
			    if (SvAMAGIC(*p1))
				overloading = 1;
			    else
				(void)sv_2nv(*p1);
			}
			if (all_SIVs && !SvSIOK(*p1))
			    all_SIVs = 0;
		    }
		}
		else {
		    if (!SvPOK(*p1)) {
			if (SvAMAGIC(*p1))
			    overloading = 1;
			else
			    (void)sv_2pv_flags(*p1, 0,
					       SV_GMAGIC|SV_CONST_RETURN);
		    }
		}
	    }
	    p1++;
	}
	else
	    max--;
    }
    if (sorting_av)
	AvFILLp(av) = max-1;

    if (max > 1) {
	SV **start;
	if (PL_sortcop) {
	    PERL_CONTEXT *cx;
	    SV** newsp;
	    const bool oldcatch = CATCH_GET;

	    SAVETMPS;
	    SAVEOP();

	    CATCH_SET(TRUE);
	    PUSHSTACKi(PERLSI_SORT);
	    if (!hasargs && !is_xsub) {
		SAVESPTR(PL_firstgv);
		SAVESPTR(PL_secondgv);
		SAVESPTR(PL_sortstash);
		PL_firstgv = gv_fetchpvs("a", GV_ADD|GV_NOTQUAL, SVt_PV);
		PL_secondgv = gv_fetchpvs("b", GV_ADD|GV_NOTQUAL, SVt_PV);
		PL_sortstash = stash;
		SAVESPTR(GvSV(PL_firstgv));
		SAVESPTR(GvSV(PL_secondgv));
	    }

	    PUSHBLOCK(cx, CXt_NULL, PL_stack_base);
	    if (!(flags & OPf_SPECIAL)) {
		cx->cx_type = CXt_SUB;
		cx->blk_gimme = G_SCALAR;
		/* If our comparison routine is already active (CvDEPTH is
		 * is not 0),  then PUSHSUB does not increase the refcount,
		 * so we have to do it ourselves, because the LEAVESUB fur-
		 * ther down lowers it. */
		if (CvDEPTH(cv)) SvREFCNT_inc_simple_void_NN(cv);
		PUSHSUB(cx);
		if (!is_xsub) {
		    AV* const padlist = CvPADLIST(cv);

		    if (++CvDEPTH(cv) >= 2) {
			PERL_STACK_OVERFLOW_CHECK();
			pad_push(padlist, CvDEPTH(cv));
		    }
		    SAVECOMPPAD();
		    PAD_SET_CUR_NOSAVE(padlist, CvDEPTH(cv));

		    if (hasargs) {
			/* This is mostly copied from pp_entersub */
			AV * const av = MUTABLE_AV(PAD_SVl(0));

			cx->blk_sub.savearray = GvAV(PL_defgv);
			GvAV(PL_defgv) = MUTABLE_AV(SvREFCNT_inc_simple(av));
			CX_CURPAD_SAVE(cx->blk_sub);
			cx->blk_sub.argarray = av;
		    }

		}
	    }
	    cx->cx_type |= CXp_MULTICALL;
	    
	    start = p1 - max;
	    sortsvp(aTHX_ start, max,
		    (is_xsub ? S_sortcv_xsub : hasargs ? S_sortcv_stacked : S_sortcv),
		    sort_flags);

	    if (!(flags & OPf_SPECIAL)) {
		LEAVESUB(cv);
		if (!is_xsub)
		    CvDEPTH(cv)--;
	    }
	    POPBLOCK(cx,PL_curpm);
	    PL_stack_sp = newsp;
	    POPSTACK;
	    CATCH_SET(oldcatch);
	}
	else {
	    MEXTEND(SP, 20);	/* Can't afford stack realloc on signal. */
	    start = sorting_av ? AvARRAY(av) : ORIGMARK+1;
	    sortsvp(aTHX_ start, max,
		    (priv & OPpSORT_NUMERIC)
		        ? ( ( ( priv & OPpSORT_INTEGER) || all_SIVs)
			    ? ( overloading ? S_amagic_i_ncmp : S_sv_i_ncmp)
			    : ( overloading ? S_amagic_ncmp : S_sv_ncmp ) )
			: ( IN_LOCALE_RUNTIME
			    ? ( overloading
				? (SVCOMPARE_t)S_amagic_cmp_locale
				: (SVCOMPARE_t)sv_cmp_locale_static)
			    : ( overloading ? (SVCOMPARE_t)S_amagic_cmp : (SVCOMPARE_t)sv_cmp_static)),
		    sort_flags);
	}
	if ((priv & OPpSORT_REVERSE) != 0) {
	    SV **q = start+max-1;
	    while (start < q) {
		SV * const tmp = *start;
		*start++ = *q;
		*q-- = tmp;
	    }
	}
    }
    if (sorting_av)
	SvREADONLY_off(av);
    else if (av && !sorting_av) {
	/* simulate pp_aassign of tied AV */
	SV** const base = MARK+1;
	for (i=0; i < max; i++) {
	    base[i] = newSVsv(base[i]);
	}
	av_clear(av);
	av_extend(av, max);
	for (i=0; i < max; i++) {
	    SV * const sv = base[i];
	    SV ** const didstore = av_store(av, i, sv);
	    if (SvSMAGICAL(sv))
		mg_set(sv);
	    if (!didstore)
		sv_2mortal(sv);
	}
    }
    LEAVE;
    PL_stack_sp = ORIGMARK + (sorting_av ? 0 : max);
    return nextop;
}

static I32
S_sortcv(pTHX_ SV *const a, SV *const b)
{
    dVAR;
    const I32 oldsaveix = PL_savestack_ix;
    const I32 oldscopeix = PL_scopestack_ix;
    I32 result;
 
    PERL_ARGS_ASSERT_SORTCV;

    GvSV(PL_firstgv) = a;
    GvSV(PL_secondgv) = b;
    PL_stack_sp = PL_stack_base;
    PL_op = PL_sortcop;
    CALLRUNOPS(aTHX);
    if (PL_stack_sp != PL_stack_base + 1)
	Perl_croak(aTHX_ "Sort subroutine didn't return single value");
    result = SvIV(*PL_stack_sp);
    while (PL_scopestack_ix > oldscopeix) {
	LEAVE;
    }
    leave_scope(oldsaveix);
    return result;
}

static I32
S_sortcv_stacked(pTHX_ SV *const a, SV *const b)
{
    dVAR;
    const I32 oldsaveix = PL_savestack_ix;
    const I32 oldscopeix = PL_scopestack_ix;
    I32 result;
    AV * const av = GvAV(PL_defgv);

    PERL_ARGS_ASSERT_SORTCV_STACKED;

    if (AvMAX(av) < 1) {
	SV** ary = AvALLOC(av);
	if (AvARRAY(av) != ary) {
	    AvMAX(av) += AvARRAY(av) - AvALLOC(av);
	    AvARRAY(av) = ary;
	}
	if (AvMAX(av) < 1) {
	    AvMAX(av) = 1;
	    Renew(ary,2,SV*);
	    AvARRAY(av) = ary;
	}
    }
    AvFILLp(av) = 1;

    AvARRAY(av)[0] = a;
    AvARRAY(av)[1] = b;
    PL_stack_sp = PL_stack_base;
    PL_op = PL_sortcop;
    CALLRUNOPS(aTHX);
    if (PL_stack_sp != PL_stack_base + 1)
	Perl_croak(aTHX_ "Sort subroutine didn't return single value");
    result = SvIV(*PL_stack_sp);
    while (PL_scopestack_ix > oldscopeix) {
	LEAVE;
    }
    leave_scope(oldsaveix);
    return result;
}

static I32
S_sortcv_xsub(pTHX_ SV *const a, SV *const b)
{
    dVAR; dSP;
    const I32 oldsaveix = PL_savestack_ix;
    const I32 oldscopeix = PL_scopestack_ix;
    CV * const cv=MUTABLE_CV(PL_sortcop);
    I32 result;

    PERL_ARGS_ASSERT_SORTCV_XSUB;

    SP = PL_stack_base;
    PUSHMARK(SP);
    EXTEND(SP, 2);
    *++SP = a;
    *++SP = b;
    PUTBACK;
    (void)(*CvXSUB(cv))(aTHX_ cv);
    if (PL_stack_sp != PL_stack_base + 1)
	Perl_croak(aTHX_ "Sort subroutine didn't return single value");
    result = SvIV(*PL_stack_sp);
    while (PL_scopestack_ix > oldscopeix) {
	LEAVE;
    }
    leave_scope(oldsaveix);
    return result;
}


static I32
S_sv_ncmp(pTHX_ SV *const a, SV *const b)
{
    const NV nv1 = SvNSIV(a);
    const NV nv2 = SvNSIV(b);

    PERL_ARGS_ASSERT_SV_NCMP;

    return nv1 < nv2 ? -1 : nv1 > nv2 ? 1 : 0;
}

static I32
S_sv_i_ncmp(pTHX_ SV *const a, SV *const b)
{
    const IV iv1 = SvIV(a);
    const IV iv2 = SvIV(b);

    PERL_ARGS_ASSERT_SV_I_NCMP;

    return iv1 < iv2 ? -1 : iv1 > iv2 ? 1 : 0;
}

#define tryCALL_AMAGICbin(left,right,meth) \
    (PL_amagic_generation && (SvAMAGIC(left)||SvAMAGIC(right))) \
	? amagic_call(left, right, CAT2(meth,_amg), 0) \
	: NULL;

#define SORT_NORMAL_RETURN_VALUE(val)  (((val) > 0) ? 1 : ((val) ? -1 : 0))

static I32
S_amagic_ncmp(pTHX_ register SV *const a, register SV *const b)
{
    dVAR;
    SV * const tmpsv = tryCALL_AMAGICbin(a,b,ncmp);

    PERL_ARGS_ASSERT_AMAGIC_NCMP;

    if (tmpsv) {
        if (SvIOK(tmpsv)) {
            const I32 i = SvIVX(tmpsv);
            return SORT_NORMAL_RETURN_VALUE(i);
        }
	else {
	    const NV d = SvNV(tmpsv);
	    return SORT_NORMAL_RETURN_VALUE(d);
	}
     }
     return S_sv_ncmp(aTHX_ a, b);
}

static I32
S_amagic_i_ncmp(pTHX_ register SV *const a, register SV *const b)
{
    dVAR;
    SV * const tmpsv = tryCALL_AMAGICbin(a,b,ncmp);

    PERL_ARGS_ASSERT_AMAGIC_I_NCMP;

    if (tmpsv) {
        if (SvIOK(tmpsv)) {
            const I32 i = SvIVX(tmpsv);
            return SORT_NORMAL_RETURN_VALUE(i);
        }
	else {
	    const NV d = SvNV(tmpsv);
	    return SORT_NORMAL_RETURN_VALUE(d);
	}
    }
    return S_sv_i_ncmp(aTHX_ a, b);
}

static I32
S_amagic_cmp(pTHX_ register SV *const str1, register SV *const str2)
{
    dVAR;
    SV * const tmpsv = tryCALL_AMAGICbin(str1,str2,scmp);

    PERL_ARGS_ASSERT_AMAGIC_CMP;

    if (tmpsv) {
        if (SvIOK(tmpsv)) {
            const I32 i = SvIVX(tmpsv);
            return SORT_NORMAL_RETURN_VALUE(i);
        }
	else {
	    const NV d = SvNV(tmpsv);
	    return SORT_NORMAL_RETURN_VALUE(d);
	}
    }
    return sv_cmp(str1, str2);
}

static I32
S_amagic_cmp_locale(pTHX_ register SV *const str1, register SV *const str2)
{
    dVAR;
    SV * const tmpsv = tryCALL_AMAGICbin(str1,str2,scmp);

    PERL_ARGS_ASSERT_AMAGIC_CMP_LOCALE;

    if (tmpsv) {
        if (SvIOK(tmpsv)) {
            const I32 i = SvIVX(tmpsv);
            return SORT_NORMAL_RETURN_VALUE(i);
        }
	else {
	    const NV d = SvNV(tmpsv);
	    return SORT_NORMAL_RETURN_VALUE(d);
	}
    }
    return sv_cmp_locale(str1, str2);
}

/*
 * Local variables:
 * c-indentation-style: bsd
 * c-basic-offset: 4
 * indent-tabs-mode: t
 * End:
 *
 * ex: set ts=8 sts=4 sw=4 noet:
 */