From 463e7df6ad9504b38960d0a7296d537792ed3e87 Mon Sep 17 00:00:00 2001
From: Klaus Post <klauspost@gmail.com>
Date: Thu, 8 Sep 2022 14:30:14 +0200
Subject: [PATCH] s2: Improve "best" compression (#658)

Allow skipping bytes in the beginning. Since we already have pretty good matches we aren't too likely to find better.

Improvement in the 0.0x%->0.10 range.
---
 s2/encode_best.go | 27 ++++++++++++++++++++++-----
 1 file changed, 22 insertions(+), 5 deletions(-)

diff --git a/s2/encode_best.go b/s2/encode_best.go
index 61eb73b2dd..1b7ea394fa 100644
--- a/s2/encode_best.go
+++ b/s2/encode_best.go
@@ -177,14 +177,21 @@ func encodeBlockBest(dst, src []byte) (d int) {
 						best = bestOf(best, matchAt(getPrev(nextLong), s, uint32(cv), false))
 					}
 					// Search for a match at best match end, see if that is better.
-					if sAt := best.s + best.length; sAt < sLimit {
-						sBack := best.s
-						backL := best.length
+					// Allow some bytes at the beginning to mismatch.
+					// Sweet spot is around 1-2 bytes, but depends on input.
+					// The skipped bytes are tested in Extend backwards,
+					// and still picked up as part of the match if they do.
+					const skipBeginning = 2
+					const skipEnd = 1
+					if sAt := best.s + best.length - skipEnd; sAt < sLimit {
+
+						sBack := best.s + skipBeginning - skipEnd
+						backL := best.length - skipBeginning
 						// Load initial values
 						cv = load64(src, sBack)
-						// Search for mismatch
+
+						// Grab candidates...
 						next := lTable[hash8(load64(src, sAt), lTableBits)]
-						//next := sTable[hash4(load64(src, sAt), sTableBits)]
 
 						if checkAt := getCur(next) - backL; checkAt > 0 {
 							best = bestOf(best, matchAt(checkAt, sBack, uint32(cv), false))
@@ -192,6 +199,16 @@ func encodeBlockBest(dst, src []byte) (d int) {
 						if checkAt := getPrev(next) - backL; checkAt > 0 {
 							best = bestOf(best, matchAt(checkAt, sBack, uint32(cv), false))
 						}
+						// Disabled: Extremely small gain
+						if false {
+							next = sTable[hash4(load64(src, sAt), sTableBits)]
+							if checkAt := getCur(next) - backL; checkAt > 0 {
+								best = bestOf(best, matchAt(checkAt, sBack, uint32(cv), false))
+							}
+							if checkAt := getPrev(next) - backL; checkAt > 0 {
+								best = bestOf(best, matchAt(checkAt, sBack, uint32(cv), false))
+							}
+						}
 					}
 				}
 			}