From 6810e521b26da3a38183dd48ea2ebd6075ce88b4 Mon Sep 17 00:00:00 2001 From: David Geier Date: Fri, 6 Nov 2026 10:51:50 +0100 Subject: [PATCH v12 2/4] Optimize trigram sort comparators CMPTRGM_SIGNED() and CMPTRGM_UNSIGNED() now assemble the trigram into a single uint32 from explicit byte loads, with trg[0] as the most significant byte, making the comparison endian-agnostic and avoiding a read past the 3-byte trgm. The signed variant flips the sign bit of each byte so that signed-char ordering maps onto unsigned ordering, in the spirit of radix sort. --- contrib/pg_trgm/trgm_op.c | 46 +++++++++++++++++++++++++++++---------- 1 file changed, 34 insertions(+), 12 deletions(-) diff --git a/contrib/pg_trgm/trgm_op.c b/contrib/pg_trgm/trgm_op.c index 22bcc3c3361..8f9b64649d4 100644 --- a/contrib/pg_trgm/trgm_op.c +++ b/contrib/pg_trgm/trgm_op.c @@ -185,30 +185,52 @@ _PG_init(void) MarkGUCPrefixReserved("pg_trgm"); } -#define CMPCHAR(a,b) ( ((a)==(b)) ? 0 : ( ((a)<(b)) ? -1 : 1 ) ) - /* * Functions for comparing two trgms while treating each char as "signed char" or * "unsigned char". + * + * Trigrams are stored in memory as trg[0], trg[1], trg[2], ordered from most + * significant trigram to least significant trigram. + * The code adheres to that by shifting trg[0] to the most significant byte, + * trg[1] to the second most significant byte anbd trg[2] to the last significant + * byte. Only loading bytes makes this code endian-agnostic. + * */ +static pg_always_inline int +cmptrgm(const trgm a, const trgm b, bool char_is_signed) +{ + const uint8 *ua = (const uint8 *)a; + const uint8 *ub = (const uint8 *)b; + uint32 ia = ((uint32)ua[0] << 16) | ((uint32)ua[1] << 8) | ua[2]; + uint32 ib = ((uint32)ub[0] << 16) | ((uint32)ub[1] << 8) | ub[2]; + + /* + * The signed comparator flips the sign bit of each of the three trigrams to + * ensure correct ordering on compilers that tread char as signed. + * + * Background: When treated as unsigned, positive numbers sort after negative + * numbers because negative numbers have the sign bit set in two's complement. + * Flipping the sign bit reverses this order. + */ + if (char_is_signed) + { + ia ^= 0x00808080; + ib ^= 0x00808080; + } + + return (ia < ib) ? -1 : ((ia > ib) ? 1 : 0); +} + static inline int CMPTRGM_SIGNED(const void *a, const void *b) { -#define CMPPCHAR_S(a,b,i) CMPCHAR( *(((const signed char*)(a))+i), *(((const signed char*)(b))+i) ) - - return CMPPCHAR_S(a, b, 0) ? CMPPCHAR_S(a, b, 0) - : (CMPPCHAR_S(a, b, 1) ? CMPPCHAR_S(a, b, 1) - : CMPPCHAR_S(a, b, 2)); + return cmptrgm(a, b, true); } static inline int CMPTRGM_UNSIGNED(const void *a, const void *b) { -#define CMPPCHAR_UNS(a,b,i) CMPCHAR( *(((const unsigned char*)(a))+i), *(((const unsigned char*)(b))+i) ) - - return CMPPCHAR_UNS(a, b, 0) ? CMPPCHAR_UNS(a, b, 0) - : (CMPPCHAR_UNS(a, b, 1) ? CMPPCHAR_UNS(a, b, 1) - : CMPPCHAR_UNS(a, b, 2)); + return cmptrgm(a, b, false); } /* -- 2.53.0