From 7c832f65e78cbc648a416af6225aafc825ce695d Mon Sep 17 00:00:00 2001
From: Sun Yimin <emmansun@users.noreply.github.com>
Date: Tue, 21 May 2024 17:38:04 +0800
Subject: [PATCH] kdf-sm3: optimize arm64

---
 sm3/sm3blocks_arm64.s | 12 ++++++------
 sm3/sm3blocks_test.go |  8 --------
 2 files changed, 6 insertions(+), 14 deletions(-)

diff --git a/sm3/sm3blocks_arm64.s b/sm3/sm3blocks_arm64.s
index 02f0303..29e0ac7 100644
--- a/sm3/sm3blocks_arm64.s
+++ b/sm3/sm3blocks_arm64.s
@@ -197,7 +197,9 @@ TEXT ·blockMultBy4(SB), NOSPLIT, $0
 	MOVD (srcPtrPtr), srcPtr4
 
 loop:
+	// reset wordPtr
 	MOVD wordStart, wordPtr
+
 	// load message block
 	prepare4Words
 	prepare4Words
@@ -271,9 +273,8 @@ loop:
 	ROUND_16_63(62, T30, c, d, e, f, g, h, a, b)
 	ROUND_16_63(63, T31, b, c, d, e, f, g, h, a)
 
-	MOVD statePtr, R20
-	VLD1.P 64(R20), [V8.S4, V9.S4, V10.S4, V11.S4]
-	VLD1 (R20), [V12.S4, V13.S4, V14.S4, V15.S4]
+	VLD1.P 64(statePtr), [V8.S4, V9.S4, V10.S4, V11.S4]
+	VLD1 (statePtr), [V12.S4, V13.S4, V14.S4, V15.S4]
 	VEOR a.B16, V8.B16, a.B16
 	VEOR b.B16, V9.B16, b.B16
 	VEOR c.B16, V10.B16, c.B16
@@ -282,9 +283,8 @@ loop:
 	VEOR f.B16, V13.B16, f.B16
 	VEOR g.B16, V14.B16, g.B16
 	VEOR h.B16, V15.B16, h.B16
-	MOVD statePtr, R20
-	VST1.P [a.S4, b.S4, c.S4, d.S4], 64(R20)
-	VST1 [e.S4, f.S4, g.S4, h.S4], (R20)
+	VST1.P [e.S4, f.S4, g.S4, h.S4], -64(statePtr)
+	VST1 [a.S4, b.S4, c.S4, d.S4], (statePtr)
 
 	SUB $1, blockCount
 	CBNZ blockCount, loop
diff --git a/sm3/sm3blocks_test.go b/sm3/sm3blocks_test.go
index 63c5bae..5e78ce3 100644
--- a/sm3/sm3blocks_test.go
+++ b/sm3/sm3blocks_test.go
@@ -3,7 +3,6 @@
 package sm3
 
 import (
-	"encoding/binary"
 	"fmt"
 	"testing"
 )
@@ -54,13 +53,6 @@ func TestBlockMultBy4(t *testing.T) {
 	buffer := make([]byte, preallocSize)
 	blockMultBy4(&digs[0], &p[0], &buffer[0], 1)
 	expected := "[66c7f0f4 62eeedd9 d1f2d46b dc10e4e2 4167c487 5cf2f7a2 297da02b 8f4ba8e0]"
-	for i := 0; i < 128+68*4*4; i += 64 {
-		fmt.Printf("%08x %08x %08x %08x ", binary.LittleEndian.Uint32(buffer[i:]), binary.LittleEndian.Uint32(buffer[i+4:]), binary.LittleEndian.Uint32(buffer[i+8:]), binary.LittleEndian.Uint32(buffer[i+12:]))
-		fmt.Printf("%08x %08x %08x %08x ", binary.LittleEndian.Uint32(buffer[i+16:]), binary.LittleEndian.Uint32(buffer[i+20:]), binary.LittleEndian.Uint32(buffer[i+24:]), binary.LittleEndian.Uint32(buffer[i+28:]))
-		fmt.Printf("%08x %08x %08x %08x ", binary.LittleEndian.Uint32(buffer[i+32:]), binary.LittleEndian.Uint32(buffer[i+36:]), binary.LittleEndian.Uint32(buffer[i+40:]), binary.LittleEndian.Uint32(buffer[i+44:]))
-		fmt.Printf("%08x %08x %08x %08x ", binary.LittleEndian.Uint32(buffer[i+48:]), binary.LittleEndian.Uint32(buffer[i+52:]), binary.LittleEndian.Uint32(buffer[i+56:]), binary.LittleEndian.Uint32(buffer[i+60:]))
-		fmt.Println()
-	}
 	s := fmt.Sprintf("%x", digs[0][:])
 	if s != expected {
 		t.Errorf("digs[0] got %s", s)