From 5847c49dcf4f98adb0d3bfc16a43fb3135c852b4 Mon Sep 17 00:00:00 2001 From: Peter Geoghegan Date: Fri, 3 Jul 2015 13:48:08 -0700 Subject: [PATCH 2/4] Use unsigned integer abbreviated keys for text Arrange for text abbreviated keys to be represented such that a simple 3-way unsigned integer comparison using the new representation is correct. This is somewhat faster on at least some platforms, since unsigned integer comparisons are now used rather than memcmp() during sorting proper. On little-endian machines, this requires a byteswap at the end of each call to conversion routine (big-endian machines use the same bitwise representation as before). --- src/backend/utils/adt/varlena.c | 20 +++++++++++--------- 1 file changed, 11 insertions(+), 9 deletions(-) diff --git a/src/backend/utils/adt/varlena.c b/src/backend/utils/adt/varlena.c index 2fbbf54..63ca437 100644 --- a/src/backend/utils/adt/varlena.c +++ b/src/backend/utils/adt/varlena.c @@ -1967,25 +1967,25 @@ done: static int bttextcmp_abbrev(Datum x, Datum y, SortSupport ssup) { - char *a = (char *) &x; - char *b = (char *) &y; - int result; - - result = memcmp(a, b, sizeof(Datum)); - /* - * When result = 0, the core system will call bttextfastcmp_c() or + * When 0 is returned, the core system will call bttextfastcmp_c() or * bttextfastcmp_locale(). Even a strcmp() on two non-truncated strxfrm() * blobs cannot indicate *equality* authoritatively, for the same reason * that there is a strcoll() tie-breaker call to strcmp() in varstr_cmp(). */ - return result; + if (x > y) + return 1; + else if (x == y) + return 0; + else + return -1; } /* * Conversion routine for sortsupport. Converts original text to abbreviated * key representation. Our encoding strategy is simple -- pack the first 8 - * bytes of a strxfrm() blob into a Datum. + * bytes of a strxfrm() blob into a Datum (on little-endian machines, the 8 + * bytes are stored in reverse order), and treat it as an unsigned integer. */ static Datum bttext_abbrev_convert(Datum original, SortSupport ssup) @@ -2104,6 +2104,8 @@ bttext_abbrev_convert(Datum original, SortSupport ssup) addHyperLogLog(&tss->abbr_card, hash); + /* Byteswap on little-endian machines: */ + res = ABBREV_STRING_UINT(res); /* Don't leak memory here */ if (PointerGetDatum(authoritative) != original) pfree(authoritative); -- 1.9.1