From 02875939fa45246140b34554c23eedccc66ba972 Mon Sep 17 00:00:00 2001
From: John Naylor <john.naylor@postgresql.org>
Date: Sun, 24 Dec 2023 15:14:46 +0700
Subject: [PATCH v11 6/8] Try simply byte-swapping on BE machines and then
 handling like LE

---
 src/include/common/hashfn_unstable.h | 36 +++++++---------------------
 1 file changed, 9 insertions(+), 27 deletions(-)

diff --git a/src/include/common/hashfn_unstable.h b/src/include/common/hashfn_unstable.h
index 4fc9edba6e..5bc1fc88ec 100644
--- a/src/include/common/hashfn_unstable.h
+++ b/src/include/common/hashfn_unstable.h
@@ -14,6 +14,7 @@ the same hashes between versions.
 #define HASHFN_UNSTABLE_H
 
 #include "port/pg_bitutils.h"
+#include "port/pg_bswap.h"
 
 /*
  * fasthash is a modification of code taken from
@@ -152,9 +153,6 @@ fasthash_accum(fasthash_state *hs, const char *k, int len)
 #define haszero64(v) \
 	(((v) - 0x0101010101010101UL) & ~(v) & 0x8080808080808080UL)
 
-#define SIM_BE 1
-#include "port/pg_bswap.h"
-
 /*
  * With an aligned pointer, we consume the string a word at a time. Loading
  * the word containing the NUL terminator cannot segfault since page boundaries
@@ -170,39 +168,27 @@ fasthash_accum_cstring_aligned(fasthash_state *hs, const char *str)
 	uint64 zero_bytes;
 
 	Assert(PointerIsAligned(start, uint64));
-	while (true)
+	for (;;)
 	{
 		uint64 chunk = *(uint64 *)buf;
-#ifdef SIM_BE
-		uint64 low_bits = 0x7F7F7F7F7F7F7F7F;
 
-		chunk = pg_bswap64(chunk); /* simulate BE */
+#ifdef WORDS_BIGENDIAN
+		/* switch to little endian, to make later calculations easier */
+		chunk = pg_bswap64(chunk);
+#endif
 
 		/*
-		 * This expression evaluates has the useful property that all bytes in the result word
-		 * that correspond to non-zero bytes in the original word have
-		 * the value 0x00, while all bytes corresponding to zero bytes have
-		 * the value 0x80.
-		 */
-		zero_bytes = ~(((chunk & low_bits) + low_bits) | chunk | low_bits);
-#else
-		/*
-		 * On little endian machines, we can use a slightly faster calculation,
+		 * With little-endian representation, we can use this calculation,
 		 * which sets bits in the first byte in the result word
 		 * that corresponds to a zero byte in the original word.
 		 * The rest of the bytes are indeterminate, so cannot be used
-		 * on big-endian machines unless we resort to a bytewise check.
+		 * on big-endian machines without either swapping or a bytewise check.
 		 */
 		zero_bytes = haszero64(chunk);
-#endif
 		if (zero_bytes)
 			break;
 
-#ifdef SIM_BE
-		hs->accum = pg_bswap64(chunk); /* not needed with real BE, because we won't need the same answer */
-#else
 		hs->accum = chunk;
-#endif
 		fasthash_combine(hs);
 		buf += FH_SIZEOF_ACCUM;
 	}
@@ -210,14 +196,10 @@ fasthash_accum_cstring_aligned(fasthash_state *hs, const char *str)
 	/*
 	 * Bytes with set bits will be 0x80, so
 	 * calculate the first occurrence of a zero byte within the input word
-	 * by counting the number of leading (on BE) or trailing (on LE)
+	 * by counting the number of trailing (for LE)
 	 * zeros and dividing the result by 8.
 	 */
-#ifdef SIM_BE
-	remainder = (63 - pg_leftmost_one_pos64(zero_bytes)) / BITS_PER_BYTE;
-#else
 	remainder = pg_rightmost_one_pos64(zero_bytes) / BITS_PER_BYTE;
-#endif
 	fasthash_accum(hs, buf, remainder);
 	buf += remainder;
 
-- 
2.43.0