47 #define G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1) \  argument
52   row4h = _mm_xor_si128(row4h, row1h); \
55   row4h = _mm_roti_epi64(row4h, -32); \
58   row3h = _mm_add_epi64(row3h, row4h); \
66 #define G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1) \  argument
71   row4h = _mm_xor_si128(row4h, row1h); \
74   row4h = _mm_roti_epi64(row4h, -16); \
77   row3h = _mm_add_epi64(row3h, row4h); \
86 #define DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
96   t0 = _mm_alignr_epi8(row4h, row4l, 8); \
97   t1 = _mm_alignr_epi8(row4l, row4h, 8); \
99   row4h = t0;
101 #define UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
111   t0 = _mm_alignr_epi8(row4l, row4h, 8); \
112   t1 = _mm_alignr_epi8(row4h, row4l, 8); \
114   row4h = t0;
117 #define DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
123   row4l = _mm_unpackhi_epi64(row4h, _mm_unpacklo_epi64(t0, t0)); \
124   row4h = _mm_unpackhi_epi64(t0, _mm_unpacklo_epi64(row4h, row4h)); \
128 #define UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
136   row4l = _mm_unpackhi_epi64(row4l, _mm_unpacklo_epi64(row4h, row4h)); \
137   row4h = _mm_unpackhi_epi64(row4h, _mm_unpacklo_epi64(t1, t1))
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);