From: John P. Linderman Date: Thu, 13 Dec 2001 09:50:58 +0000 (-0500) Subject: (LONG!) Sort benchmarks and patch X-Git-Url: http://git.shadowcat.co.uk/gitweb/gitweb.cgi?a=commitdiff_plain;h=957d8989ea0b51caca5f3244eacb3b0e48696fe9;p=p5sagit%2Fp5-mst-13.2.git (LONG!) Sort benchmarks and patch Message-Id: <200112131450.JAA84836@raptor.research.att.com> p4raw-id: //depot/perl@13673 --- diff --git a/pp_sort.c b/pp_sort.c index 797cb22..fa76c3e 100644 --- a/pp_sort.c +++ b/pp_sort.c @@ -52,15 +52,6 @@ static I32 amagic_cmp_locale(pTHX_ SV *a, SV *b); * */ -#ifdef TESTHARNESS -#include -typedef void SV; -#define pTHX_ -#define STATIC -#define New(ID,VAR,N,TYPE) VAR=(TYPE *)malloc((N)*sizeof(TYPE)) -#define Safefree(VAR) free(VAR) -typedef int (*SVCOMPARE_t) (pTHX_ SV*, SV*); -#endif /* TESTHARNESS */ typedef char * aptr; /* pointer for arithmetic on sizes */ typedef SV * gptr; /* pointers in our lists */ @@ -181,13 +172,14 @@ typedef SV * gptr; /* pointers in our lists */ */ -static void +static IV dynprep(pTHX_ gptr *list1, gptr *list2, size_t nmemb, SVCOMPARE_t cmp) { - int sense; + I32 sense; register gptr *b, *p, *q, *t, *p2; register gptr c, *last, *r; gptr *savep; + IV runs = 0; b = list1; last = PINDEX(b, nmemb); @@ -225,7 +217,7 @@ dynprep(pTHX_ gptr *list1, gptr *list2, size_t nmemb, SVCOMPARE_t cmp) ((t + 1) == last) && ((cmp(aTHX_ *(p-1), *p) > 0) == sense)) savep = r = p = q = last; - p2 = NEXT(p2) = p2 + (p - b); + p2 = NEXT(p2) = p2 + (p - b); ++runs; if (sense) while (b < --p) { c = *b; *b++ = *p; @@ -234,7 +226,7 @@ dynprep(pTHX_ gptr *list1, gptr *list2, size_t nmemb, SVCOMPARE_t cmp) p = savep; } while (q < p) { /* simple pairs */ - p2 = NEXT(p2) = p2 + 2; + p2 = NEXT(p2) = p2 + 2; ++runs; if (sense) { c = *q++; *(q-1) = *q; @@ -242,14 +234,14 @@ dynprep(pTHX_ gptr *list1, gptr *list2, size_t nmemb, SVCOMPARE_t cmp) } else q += 2; } if (((b = p) == t) && ((t+1) == last)) { - NEXT(p2) = p2 + 1; + NEXT(p2) = p2 + 1; ++runs; b++; } q = r; } while (b < t); sense = !sense; } - return; + return runs; } @@ -405,6 +397,296 @@ S_mergesortsv(pTHX_ gptr *list1, size_t nmemb, SVCOMPARE_t cmp) 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, + * qsort on many platforms, but slower than qsort, conspicuously so, + * on others. The most likely explanation is 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 advanatge of them. + * + * By contrast, consider how the quicksort algorithm above works. + * 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. + * 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 { + 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 void +S_cfmergesortsv(pTHX_ gptr *base, size_t nmemb, SVCOMPARE_t cmp) +{ + IV i, run, runs, offset; + I32 sense, level; + int iwhich; + register gptr *f1, *f2, *t, *b, *p, *tp2, *l1, *l2, *q; + gptr *aux, *list1, *list2; + gptr *p1; + gptr small[SMALLSORT]; + gptr *which[3]; + off_runs stack[60], *stackp; + + if (nmemb <= 1) return; /* sorted trivially */ + if (nmemb <= SMALLSORT) aux = small; /* use stack for aux array */ + else { New(799,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. + */ + if ((runs = stackp->runs) == 0) { + iwhich = level & 1; + list1 = which[iwhich]; /* area where runs are now */ + list2 = which[++iwhich]; /* area for merged runs */ + do { + 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. + */ + + + 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 */ + return; +} + /* * The quicksort implementation was derived from source code contributed * by Tom Horsley. @@ -1266,7 +1548,7 @@ Perl_sortsv(pTHX_ SV **array, size_t nmemb, SVCOMPARE_t cmp) sortsvp = S_qsortsv; else { if (hints & HINT_SORT_MERGESORT) - sortsvp = S_mergesortsv; + sortsvp = S_cfmergesortsv; else sortsvp = S_mergesortsv; } @@ -1658,5 +1940,3 @@ amagic_cmp_locale(pTHX_ register SV *str1, register SV *str2) } return sv_cmp_locale(str1, str2); } - -