diff --git a/sm3/sm3blocks_arm64.s b/sm3/sm3blocks_arm64.s index 03743d3..4710db2 100644 --- a/sm3/sm3blocks_arm64.s +++ b/sm3/sm3blocks_arm64.s @@ -52,10 +52,10 @@ VLD1.P 16(srcPtr3), [V14.B16] \ VLD1.P 16(srcPtr4), [V15.B16] \ TRANSPOSE_MATRIX(V12, V13, V14, V15, tmp1, tmp2, tmp3, tmp4); \ - VREV32 V12, V12; \ - VREV32 V13, V13; \ - VREV32 V14, V14; \ - VREV32 V15, V15; \ + VREV32 V12.B16, V12.B16; \ + VREV32 V13.B16, V13.B16; \ + VREV32 V14.B16, V14.B16; \ + VREV32 V15.B16, V15.B16; \ VST1.P [V12.B16, V13.B16, V14.B16, V15.B16], 64(wordPtr) #define LOAD_T(const, T) \