diff --git a/sm3/sm3blocks_arm64.s b/sm3/sm3blocks_arm64.s
index 89bdbc9..bf93f14 100644
--- a/sm3/sm3blocks_arm64.s
+++ b/sm3/sm3blocks_arm64.s
@@ -59,7 +59,7 @@
 	VST1.P [V12.B16, V13.B16, V14.B16, V15.B16], 64(wordPtr)
 
 #define LOAD_T(const, T) \
-	MOVD $const, R20     \
+	MOVW $const, R20     \
 	VDUP R20, T.S4       \
 
 #define ROUND_00_11(index, const, a, b, c, d, e, f, g, h) \
diff --git a/sm3/sm3blocks_test.go b/sm3/sm3blocks_test.go
index 2311304..584da03 100644
--- a/sm3/sm3blocks_test.go
+++ b/sm3/sm3blocks_test.go
@@ -51,7 +51,7 @@ func createTwoBlocksBy4() [4]*byte {
 func TestBlockMultBy4(t *testing.T) {
 	digs := initState4()
 	p := createOneBlockBy4()
-	buffer := make([]byte, 1216)
+	buffer := make([]byte, preallocSize)
 	blockMultBy4(&digs[0], &p[0], &buffer[0], 1)
 	expected := "[66c7f0f4 62eeedd9 d1f2d46b dc10e4e2 4167c487 5cf2f7a2 297da02b 8f4ba8e0]"
 	for i := 128; i < 128+68*4*4; i += 64 {