8 #define load_8way(src, x0, x1, x2, x3, x4, x5, x6, x7) \  argument
9 	vmovdqu (0*16)(src), x0; \
10 	vmovdqu (1*16)(src), x1; \
11 	vmovdqu (2*16)(src), x2; \
12 	vmovdqu (3*16)(src), x3; \
13 	vmovdqu (4*16)(src), x4; \
14 	vmovdqu (5*16)(src), x5; \
15 	vmovdqu (6*16)(src), x6; \
16 	vmovdqu (7*16)(src), x7;
28 #define store_cbc_8way(src, dst, x0, x1, x2, x3, x4, x5, x6, x7) \  argument
29 	vpxor (0*16)(src), x1, x1; \
30 	vpxor (1*16)(src), x2, x2; \
31 	vpxor (2*16)(src), x3, x3; \
32 	vpxor (3*16)(src), x4, x4; \
33 	vpxor (4*16)(src), x5, x5; \
34 	vpxor (5*16)(src), x6, x6; \
35 	vpxor (6*16)(src), x7, x7; \
72 #define store_ctr_8way(src, dst, x0, x1, x2, x3, x4, x5, x6, x7) \  argument
73 	vpxor (0*16)(src), x0, x0; \
74 	vpxor (1*16)(src), x1, x1; \
75 	vpxor (2*16)(src), x2, x2; \
76 	vpxor (3*16)(src), x3, x3; \
77 	vpxor (4*16)(src), x4, x4; \
78 	vpxor (5*16)(src), x5, x5; \
79 	vpxor (6*16)(src), x6, x6; \
80 	vpxor (7*16)(src), x7, x7; \
90 #define load_xts_8way(iv, src, dst, x0, x1, x2, x3, x4, x5, x6, x7, tiv, t0, \  argument
96 	vpxor (0*16)(src), tiv, x0; \
101 	vpxor (1*16)(src), tiv, x1; \
105 	vpxor (2*16)(src), tiv, x2; \
109 	vpxor (3*16)(src), tiv, x3; \
113 	vpxor (4*16)(src), tiv, x4; \
117 	vpxor (5*16)(src), tiv, x5; \
121 	vpxor (6*16)(src), tiv, x6; \
125 	vpxor (7*16)(src), tiv, x7; \