From 86bd80e453103d01b05bb2bd8476750192dcd114 Mon Sep 17 00:00:00 2001
From: Runxi Yu
Date: Wed, 19 Nov 2025 08:00:00 +0800
Subject: adler32: Unroll update loop

---
 internal/adler32/adler32.go | 77 ++++++++++++++++++++++++++++++++++++++++-----
 1 file changed, 70 insertions(+), 7 deletions(-)

(limited to 'internal')

diff --git a/internal/adler32/adler32.go b/internal/adler32/adler32.go
index c6179789..f8a289f5 100644
--- a/internal/adler32/adler32.go
+++ b/internal/adler32/adler32.go
@@ -16,7 +16,8 @@ package adler32
 import (
 	"errors"
 	"hash"
-	"internal/byteorder"
+
+	"git.sr.ht/~runxiyu/furgit/internal/byteorder"
 )
 
 const (
@@ -86,26 +87,88 @@ func (d *digest) Clone() (hash.Cloner, error) {
 // Add p to the running checksum d.
 func update(d digest, p []byte) digest {
 	s1, s2 := uint32(d&0xffff), uint32(d>>16)
+
 	for len(p) > 0 {
 		var q []byte
 		if len(p) > nmax {
 			p, q = p[:nmax], p[nmax:]
 		}
-		for len(p) >= 4 {
-			s1 += uint32(p[0])
+
+		for len(p) >= 32 {
+			v := p[:32]
+			p = p[32:]
+
+			s1 += uint32(v[0])
+			s2 += s1
+			s1 += uint32(v[1])
+			s2 += s1
+			s1 += uint32(v[2])
+			s2 += s1
+			s1 += uint32(v[3])
+			s2 += s1
+			s1 += uint32(v[4])
+			s2 += s1
+			s1 += uint32(v[5])
+			s2 += s1
+			s1 += uint32(v[6])
+			s2 += s1
+			s1 += uint32(v[7])
+			s2 += s1
+			s1 += uint32(v[8])
+			s2 += s1
+			s1 += uint32(v[9])
+			s2 += s1
+			s1 += uint32(v[10])
+			s2 += s1
+			s1 += uint32(v[11])
+			s2 += s1
+			s1 += uint32(v[12])
+			s2 += s1
+			s1 += uint32(v[13])
 			s2 += s1
-			s1 += uint32(p[1])
+			s1 += uint32(v[14])
 			s2 += s1
-			s1 += uint32(p[2])
+			s1 += uint32(v[15])
 			s2 += s1
-			s1 += uint32(p[3])
+			s1 += uint32(v[16])
+			s2 += s1
+			s1 += uint32(v[17])
+			s2 += s1
+			s1 += uint32(v[18])
+			s2 += s1
+			s1 += uint32(v[19])
+			s2 += s1
+			s1 += uint32(v[20])
+			s2 += s1
+			s1 += uint32(v[21])
+			s2 += s1
+			s1 += uint32(v[22])
+			s2 += s1
+			s1 += uint32(v[23])
+			s2 += s1
+			s1 += uint32(v[24])
+			s2 += s1
+			s1 += uint32(v[25])
+			s2 += s1
+			s1 += uint32(v[26])
+			s2 += s1
+			s1 += uint32(v[27])
+			s2 += s1
+			s1 += uint32(v[28])
+			s2 += s1
+			s1 += uint32(v[29])
+			s2 += s1
+			s1 += uint32(v[30])
+			s2 += s1
+			s1 += uint32(v[31])
 			s2 += s1
-			p = p[4:]
 		}
+
 		for _, x := range p {
 			s1 += uint32(x)
 			s2 += s1
 		}
+
 		s1 %= mod
 		s2 %= mod
 		p = q
-- 
cgit v1.3.1-10-gc9f91