46 #define G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1) \  argument
47   row1l = _mm_add_epi64(_mm_add_epi64(row1l, b0), row2l); \
50   row4l = _mm_xor_si128(row4l, row1l); \
65 #define G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1) \  argument
66   row1l = _mm_add_epi64(_mm_add_epi64(row1l, b0), row2l); \
69   row4l = _mm_xor_si128(row4l, row1l); \
85 #define DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
100 #define UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
116 #define DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
127 #define UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
148   G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
150   G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
151   DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h); \
153   G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
155   G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
156   UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h);