46 #define G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1) \  argument
51   row4h = _mm_xor_si128(row4h, row1h); \
54   row4h = _mm_roti_epi64(row4h, -32); \
57   row3h = _mm_add_epi64(row3h, row4h); \
65 #define G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1) \  argument
70   row4h = _mm_xor_si128(row4h, row1h); \
73   row4h = _mm_roti_epi64(row4h, -16); \
76   row3h = _mm_add_epi64(row3h, row4h); \
85 #define DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
95   t0 = _mm_alignr_epi8(row4h, row4l, 8); \
96   t1 = _mm_alignr_epi8(row4l, row4h, 8); \
98   row4h = t0;
100 #define UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
110   t0 = _mm_alignr_epi8(row4l, row4h, 8); \
111   t1 = _mm_alignr_epi8(row4h, row4l, 8); \
113   row4h = t0;
116 #define DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
122   row4l = _mm_unpackhi_epi64(row4h, _mm_unpacklo_epi64(t0, t0)); \
123   row4h = _mm_unpackhi_epi64(t0, _mm_unpacklo_epi64(row4h, row4h)); \
127 #define UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
135   row4l = _mm_unpackhi_epi64(row4l, _mm_unpacklo_epi64(row4h, row4h)); \
136   row4h = _mm_unpackhi_epi64(row4h, _mm_unpacklo_epi64(t1, t1))
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);