47 #define G1(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1) \  argument
51   row4l = _mm_xor_si128(row4l, row1l); \
54   row4l = _mm_roti_epi64(row4l, -32); \
57   row3l = _mm_add_epi64(row3l, row4l); \
66 #define G2(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h,b0,b1) \  argument
70   row4l = _mm_xor_si128(row4l, row1l); \
73   row4l = _mm_roti_epi64(row4l, -16); \
76   row3l = _mm_add_epi64(row3l, row4l); \
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); \
98   row4l = t1; \
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); \
113   row4l = t1; \
117 #define DIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
118   t0 = row4l;\
120   row4l = row3l;\
122   row3h = row4l;\
123   row4l = _mm_unpackhi_epi64(row4h, _mm_unpacklo_epi64(t0, t0)); \
128 #define UNDIAGONALIZE(row1l,row2l,row3l,row4l,row1h,row2h,row3h,row4h) \  argument
133   t1 = row4l;\
136   row4l = _mm_unpackhi_epi64(row4l, _mm_unpacklo_epi64(row4h, row4h)); \
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);