X-Git-Url: http://git.shadowcat.co.uk/gitweb/gitweb.cgi?a=blobdiff_plain;f=pp_sort.c;h=8d7f48f752c9b2f16aa7503c3fcc56e98d0977ae;hb=b5f8cc5c1ad883dce8b5a96bed64f2340aa86716;hp=fa76c3e57b132e7db82c2c4a0aa02fa4da334d8f;hpb=957d8989ea0b51caca5f3244eacb3b0e48696fe9;p=p5sagit%2Fp5-mst-13.2.git diff --git a/pp_sort.c b/pp_sort.c index fa76c3e..8d7f48f 100644 --- a/pp_sort.c +++ b/pp_sort.c @@ -1,6 +1,7 @@ /* pp_sort.c * - * Copyright (c) 1991-2001, Larry Wall + * Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999, + * 2000, 2001, 2002, 2003, 2004, 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. @@ -16,6 +17,11 @@ #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 + static I32 sortcv(pTHX_ SV *a, SV *b); static I32 sortcv_stacked(pTHX_ SV *a, SV *b); static I32 sortcv_xsub(pTHX_ SV *a, SV *b); @@ -29,10 +35,9 @@ static I32 amagic_cmp_locale(pTHX_ SV *a, SV *b); #define sv_cmp_static Perl_sv_cmp #define sv_cmp_locale_static Perl_sv_cmp_locale -#define SORTHINTS(hintsvp) \ - ((PL_hintgv && \ - (hintsvp = hv_fetch(GvHV(PL_hintgv), "SORT", 4, FALSE))) ? \ - (I32)SvIV(*hintsvp) : 0) +#define SORTHINTS(hintsv) \ + (((hintsv) = GvSV(gv_fetchpv("sort::hints", GV_ADDMULTI, SVt_IV))), \ + (SvIOK(hintsv) ? ((I32)SvIV(hintsv)) : 0)) #ifndef SMALLSORT #define SMALLSORT (200) @@ -245,164 +250,9 @@ dynprep(pTHX_ gptr *list1, gptr *list2, size_t nmemb, SVCOMPARE_t cmp) } -/* Overview of bmerge variables: -** -** list1 and list2 address the main and auxiliary arrays. -** They swap identities after each merge pass. -** Base points to the original list1, so we can tell if -** the pointers ended up where they belonged (or must be copied). -** -** When we are merging two lists, f1 and f2 are the next elements -** on the respective lists. l1 and l2 mark the end of the lists. -** tp2 is the current location in the merged list. -** -** p1 records where f1 started. -** After the merge, a new descriptor is built there. -** -** p2 is a ``parallel'' pointer in (what starts as) descriptor space. -** It is used to identify and delimit the runs. -** -** In the heat of determining where q, the greater of the f1/f2 elements, -** belongs in the other list, b, t and p, represent bottom, top and probe -** locations, respectively, in the other list. -** They make convenient temporary pointers in other places. -*/ - -STATIC void -S_mergesortsv(pTHX_ gptr *list1, size_t nmemb, SVCOMPARE_t cmp) -{ - int i, run; - int sense; - register gptr *f1, *f2, *t, *b, *p, *tp2, *l1, *l2, *q; - gptr *aux, *list2, *p2, *last; - gptr *base = list1; - gptr *p1; - gptr small[SMALLSORT]; - - if (nmemb <= 1) return; /* sorted trivially */ - if (nmemb <= SMALLSORT) list2 = small; /* use stack for aux array */ - else { New(799,list2,nmemb,gptr); } /* allocate auxilliary array */ - aux = list2; - dynprep(aTHX_ list1, list2, nmemb, cmp); - last = PINDEX(list2, nmemb); - while (NEXT(list2) != last) { - /* More than one run remains. Do some merging to reduce runs. */ - l2 = p1 = list1; - for (tp2 = p2 = list2; p2 != last;) { - /* The new first run begins where the old second list ended. - ** Use the p2 ``parallel'' pointer to identify the end of the run. - */ - f1 = l2; - t = NEXT(p2); - f2 = l1 = POTHER(t, list2, list1); - if (t != last) t = NEXT(t); - l2 = POTHER(t, list2, list1); - p2 = 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. - */ - - - 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); - } - t = list1; - list1 = list2; - list2 = t; - last = PINDEX(list2, nmemb); - } - if (base == list2) { - last = PINDEX(list1, nmemb); - FROMTOUPTO(list1, list2, last); - } - if (aux != small) Safefree(aux); /* free iff allocated */ - return; -} - - -/* What perl needs (least) is another sort implementation in the core. - * So what's the story? The short (by jpl's standards) story is that - * the merge sort above, in use since 5.7, is as fast as, or faster than, +/* 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 is platform-specific + * 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 @@ -411,7 +261,7 @@ S_mergesortsv(pTHX_ gptr *list1, size_t nmemb, SVCOMPARE_t cmp) * many levels of cache exist, quicksort will "find" them, and, * as long as smaller is faster, take advanatge of them. * - * By contrast, consider how the quicksort algorithm above works. + * By contrast, consider how the original mergesort algorithm worked. * Suppose we have five runs (each typically of length 2 after dynprep). * * pass base aux @@ -478,9 +328,6 @@ S_mergesortsv(pTHX_ gptr *list1, size_t nmemb, SVCOMPARE_t cmp) * 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. - * If it's as good as the original mergesort implementation on all - * platforms, it should replace that implementation. For benchmarking, - * though, it is convenient to have both implementations available. */ typedef struct { @@ -489,7 +336,7 @@ typedef struct { } off_runs; /* pseudo-stack element */ STATIC void -S_cfmergesortsv(pTHX_ gptr *base, size_t nmemb, SVCOMPARE_t cmp) +S_mergesortsv(pTHX_ gptr *base, size_t nmemb, SVCOMPARE_t cmp) { IV i, run, runs, offset; I32 sense, level; @@ -911,7 +758,7 @@ S_qsortsvu(pTHX_ SV ** array, size_t num_elts, SVCOMPARE_t compare) register size_t n, j; register SV **q; for (n = num_elts, q = array; n > 1; ) { - j = n-- * Drand01(); + j = (size_t)(n-- * Drand01()); temp = q[j]; q[j] = q[n]; q[n] = temp; @@ -1440,7 +1287,6 @@ S_qsortsvu(pTHX_ SV ** array, size_t num_elts, SVCOMPARE_t compare) * dictated by the indirect array. */ -static SVCOMPARE_t RealCmp; static I32 cmpindir(pTHX_ gptr a, gptr b) @@ -1449,7 +1295,7 @@ cmpindir(pTHX_ gptr a, gptr b) gptr *ap = (gptr *)a; gptr *bp = (gptr *)b; - if ((sense = RealCmp(aTHX_ *ap, *bp)) == 0) + if ((sense = PL_sort_RealCmp(aTHX_ *ap, *bp)) == 0) sense = (ap > bp) ? 1 : ((ap < bp) ? -1 : 0); return sense; } @@ -1457,9 +1303,9 @@ cmpindir(pTHX_ gptr a, gptr b) STATIC void S_qsortsv(pTHX_ gptr *list1, size_t nmemb, SVCOMPARE_t cmp) { - SV **hintsvp; + SV *hintsv; - if (SORTHINTS(hintsvp) & HINT_SORT_STABLE) { + if (SORTHINTS(hintsv) & HINT_SORT_STABLE) { register gptr **pp, *q; register size_t n, j, i; gptr *small[SMALLSORT], **indir, tmp; @@ -1473,8 +1319,8 @@ S_qsortsv(pTHX_ gptr *list1, size_t nmemb, SVCOMPARE_t cmp) /* Copy pointers to original array elements into indirect array */ for (n = nmemb, pp = indir, q = list1; n--; ) *pp++ = q++; - savecmp = RealCmp; /* Save current comparison routine, if any */ - RealCmp = cmp; /* Put comparison routine where cmpindir can find it */ + 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 */ S_qsortsvu(aTHX_ (gptr *)indir, nmemb, cmpindir); @@ -1519,19 +1365,23 @@ S_qsortsv(pTHX_ gptr *list1, size_t nmemb, SVCOMPARE_t cmp) /* free iff allocated */ if (indir != small) { Safefree(indir); } /* restore prevailing comparison routine */ - RealCmp = savecmp; + PL_sort_RealCmp = savecmp; } else { S_qsortsvu(aTHX_ 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); +See lib/sort.pm for details about controlling the sorting algorithm. + =cut */ @@ -1540,18 +1390,21 @@ Perl_sortsv(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp) { void (*sortsvp)(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp) = S_mergesortsv; - SV **hintsvp; + SV *hintsv; I32 hints; - if ((hints = SORTHINTS(hintsvp))) { - if (hints & HINT_SORT_QUICKSORT) - sortsvp = S_qsortsv; - else { - if (hints & HINT_SORT_MERGESORT) - sortsvp = S_cfmergesortsv; - else - sortsvp = S_mergesortsv; - } + /* Sun's Compiler (cc: WorkShop Compilers 4.2 30 Oct 1996 C 4.2) used + to miscompile this function under optimization -O. If you get test + errors related to picking the correct sort() function, try recompiling + this file without optimiziation. -- A.D. 4/2002. + */ + hints = SORTHINTS(hintsv); + if (hints & HINT_SORT_QUICKSORT) { + sortsvp = S_qsortsv; + } + else { + /* The default as of 5.8.0 is mergesort */ + sortsvp = S_mergesortsv; } sortsvp(aTHX_ array, nmemb, cmp); @@ -1560,9 +1413,9 @@ Perl_sortsv(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp) PP(pp_sort) { dSP; dMARK; dORIGMARK; - register SV **up; - SV **myorigmark = ORIGMARK; - register I32 max; + register SV **p1 = ORIGMARK+1, **p2; + register I32 max, i; + AV* av = Nullav; HV *stash; GV *gv; CV *cv = 0; @@ -1571,6 +1424,7 @@ PP(pp_sort) I32 overloading = 0; bool hasargs = FALSE; I32 is_xsub = 0; + I32 sorting_av = 0; if (gimme != G_ARRAY) { SP = MARK; @@ -1603,8 +1457,8 @@ PP(pp_sort) else if (gv) { SV *tmpstr = sv_newmortal(); gv_efullname3(tmpstr, gv, Nullch); - DIE(aTHX_ "Undefined sort subroutine \"%s\" called", - SvPVX(tmpstr)); + DIE(aTHX_ "Undefined sort subroutine \"%"SVf"\" called", + tmpstr); } else { DIE(aTHX_ "Undefined subroutine in sort"); @@ -1618,8 +1472,7 @@ PP(pp_sort) SAVEVPTR(CvROOT(cv)->op_ppaddr); CvROOT(cv)->op_ppaddr = PL_ppaddr[OP_NULL]; - SAVEVPTR(PL_curpad); - PL_curpad = AvARRAY((AV*)AvARRAY(CvPADLIST(cv))[1]); + PAD_SET_CUR(CvPADLIST(cv), 1); } } } @@ -1628,24 +1481,54 @@ PP(pp_sort) stash = CopSTASH(PL_curcop); } - up = myorigmark + 1; - while (MARK < SP) { /* This may or may not shift down one here. */ - /*SUPPRESS 560*/ - if ((*up = *++MARK)) { /* Weed out nulls. */ - SvTEMP_off(*up); - if (!PL_sortcop && !SvPOK(*up)) { + /* 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 (PL_op->op_private & OPpSORT_INPLACE) { + assert( MARK+1 == SP && *SP && SvTYPE(*SP) == SVt_PVAV); + (void)POPMARK; /* remove mark associated with ex-OP_AASSIGN */ + av = (AV*)(*SP); + max = AvFILL(av) + 1; + if (SvMAGICAL(av)) { + MEXTEND(SP, max); + p2 = SP; + for (i=0; i < (U32)max; i++) { + SV **svp = av_fetch(av, i, FALSE); + *SP++ = (svp) ? *svp : Nullsv; + } + } + else { + 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 any args */ + for (i=max; i > 0 ; i--) { + if ((*p1 = *p2++)) { /* Weed out nulls. */ + SvTEMP_off(*p1); + if (!PL_sortcop && !SvPOK(*p1)) { STRLEN n_a; - if (SvAMAGIC(*up)) + if (SvAMAGIC(*p1)) overloading = 1; else - (void)sv_2pv(*up, &n_a); + (void)sv_2pv(*p1, &n_a); } - up++; + p1++; } + else + max--; } - max = --up - myorigmark; - if (PL_sortcop) { - if (max > 1) { + if (sorting_av) + AvFILLp(av) = max-1; + + if (max > 1) { + if (PL_sortcop) { PERL_CONTEXT *cx; SV** newsp; bool oldcatch = CATCH_GET; @@ -1663,10 +1546,6 @@ PP(pp_sort) PL_secondgv = gv_fetchpv("b", TRUE, SVt_PV); PL_sortstash = stash; } -#ifdef USE_5005THREADS - sv_lock((SV *)PL_firstgv); - sv_lock((SV *)PL_secondgv); -#endif SAVESPTR(GvSV(PL_firstgv)); SAVESPTR(GvSV(PL_secondgv)); } @@ -1676,23 +1555,19 @@ PP(pp_sort) cx->cx_type = CXt_SUB; cx->blk_gimme = G_SCALAR; PUSHSUB(cx); - if (!CvDEPTH(cv)) - (void)SvREFCNT_inc(cv); /* in preparation for POPSUB */ } PL_sortcxix = cxstack_ix; if (hasargs && !is_xsub) { /* This is mostly copied from pp_entersub */ - AV *av = (AV*)PL_curpad[0]; + AV *av = (AV*)PAD_SVl(0); -#ifndef USE_5005THREADS cx->blk_sub.savearray = GvAV(PL_defgv); GvAV(PL_defgv) = (AV*)SvREFCNT_inc(av); -#endif /* USE_5005THREADS */ - cx->blk_sub.oldcurpad = PL_curpad; + CX_CURPAD_SAVE(cx->blk_sub); cx->blk_sub.argarray = av; } - sortsv((myorigmark+1), max, + sortsv(p1-max, max, is_xsub ? sortcv_xsub : hasargs ? sortcv_stacked : sortcv); POPBLOCK(cx,PL_curpm); @@ -1700,11 +1575,9 @@ PP(pp_sort) POPSTACK; CATCH_SET(oldcatch); } - } - else { - if (max > 1) { + else { MEXTEND(SP, 20); /* Can't afford stack realloc on signal. */ - sortsv(ORIGMARK+1, max, + sortsv(sorting_av ? AvARRAY(av) : ORIGMARK+1, max, (PL_op->op_private & OPpSORT_NUMERIC) ? ( (PL_op->op_private & OPpSORT_INTEGER) ? ( overloading ? amagic_i_ncmp : sv_i_ncmp) @@ -1715,8 +1588,8 @@ PP(pp_sort) : sv_cmp_locale_static) : ( overloading ? amagic_cmp : sv_cmp_static))); if (PL_op->op_private & OPpSORT_REVERSE) { - SV **p = ORIGMARK+1; - SV **q = ORIGMARK+max; + SV **p = sorting_av ? AvARRAY(av) : ORIGMARK+1; + SV **q = p+max-1; while (p < q) { SV *tmp = *p; *p++ = *q; @@ -1725,8 +1598,28 @@ PP(pp_sort) } } } + if (av && !sorting_av) { + /* simulate pp_aassign of tied AV */ + SV *sv; + SV** base, **didstore; + for (base = ORIGMARK+1, i=0; i < max; i++) { + sv = NEWSV(28,0); + sv_setsv(sv, base[i]); + base[i] = sv; + } + av_clear(av); + av_extend(av, max); + for (i=0; i < max; i++) { + sv = base[i]; + didstore = av_store(av, i, sv); + if (SvSMAGICAL(sv)) + mg_set(sv); + if (!didstore) + sv_2mortal(sv); + } + } LEAVE; - PL_stack_sp = ORIGMARK + max; + PL_stack_sp = ORIGMARK + (sorting_av ? 0 : max); return nextop; } @@ -1761,11 +1654,7 @@ sortcv_stacked(pTHX_ SV *a, SV *b) I32 result; AV *av; -#ifdef USE_5005THREADS - av = (AV*)PL_curpad[0]; -#else av = GvAV(PL_defgv); -#endif if (AvMAX(av) < 1) { SV** ary = AvALLOC(av);