47 #define G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1) \  argument
48   row1l = _mm_add_epi64(_mm_add_epi64(row1l, b0), row2l); \
51   row4l = _mm_xor_si128(row4l, row1l); \
66 #define G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1) \  argument
67   row1l = _mm_add_epi64(_mm_add_epi64(row1l, b0), row2l); \
70   row4l = _mm_xor_si128(row4l, row1l); \
86 #define DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
101 #define UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
117 #define DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
128 #define UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
149   G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
151   G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
152   DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h); \
154   G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
156   G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1); \
157   UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h);