46 #define G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1) \  argument
48   row1h = _mm_add_epi64(_mm_add_epi64(row1h, b1), row2h); \
60   row2h = _mm_xor_si128(row2h, row3h); \
63   row2h = _mm_roti_epi64(row2h, -24); \
65 #define G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1) \  argument
67   row1h = _mm_add_epi64(_mm_add_epi64(row1h, b1), row2h); \
79   row2h = _mm_xor_si128(row2h, row3h); \
82   row2h = _mm_roti_epi64(row2h, -63); \
85 #define DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
86   t0 = _mm_alignr_epi8(row2h, row2l, 8); \
87   t1 = _mm_alignr_epi8(row2l, row2h, 8); \
89   row2h = t1; \
100 #define UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
101   t0 = _mm_alignr_epi8(row2l, row2h, 8); \
102   t1 = _mm_alignr_epi8(row2h, row2l, 8); \
104   row2h = t1; \
116 #define DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
124   row2l = _mm_unpackhi_epi64(row2l, _mm_unpacklo_epi64(row2h, row2h)); \
125   row2h = _mm_unpackhi_epi64(row2h, _mm_unpacklo_epi64(t1, t1))
127 #define UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
133   row2l = _mm_unpackhi_epi64(row2h, _mm_unpacklo_epi64(row2l, row2l)); \
134   row2h = _mm_unpackhi_epi64(t0, _mm_unpacklo_epi64(row2h, row2h)); \
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);