13 #define load_16way(src, x0, x1, x2, x3, x4, x5, x6, x7) \  argument
14 	vmovdqu (0*32)(src), x0; \
15 	vmovdqu (1*32)(src), x1; \
16 	vmovdqu (2*32)(src), x2; \
17 	vmovdqu (3*32)(src), x3; \
18 	vmovdqu (4*32)(src), x4; \
19 	vmovdqu (5*32)(src), x5; \
20 	vmovdqu (6*32)(src), x6; \
21 	vmovdqu (7*32)(src), x7;
33 #define store_cbc_16way(src, dst, x0, x1, x2, x3, x4, x5, x6, x7, t0) \  argument
35 	vinserti128 $1, (src), t0, t0; \
37 	vpxor (0*32+16)(src), x1, x1; \
38 	vpxor (1*32+16)(src), x2, x2; \
39 	vpxor (2*32+16)(src), x3, x3; \
40 	vpxor (3*32+16)(src), x4, x4; \
41 	vpxor (4*32+16)(src), x5, x5; \
42 	vpxor (5*32+16)(src), x6, x6; \
43 	vpxor (6*32+16)(src), x7, x7; \
93 #define store_ctr_16way(src, dst, x0, x1, x2, x3, x4, x5, x6, x7) \  argument
94 	vpxor (0*32)(src), x0, x0; \
95 	vpxor (1*32)(src), x1, x1; \
96 	vpxor (2*32)(src), x2, x2; \
97 	vpxor (3*32)(src), x3, x3; \
98 	vpxor (4*32)(src), x4, x4; \
99 	vpxor (5*32)(src), x5, x5; \
100 	vpxor (6*32)(src), x6, x6; \
101 	vpxor (7*32)(src), x7, x7; \
123 #define load_xts_16way(iv, src, dst, x0, x1, x2, x3, x4, x5, x6, x7, tiv, \  argument
135 	vpxor (0*32)(src), tiv, x0; \
140 	vpxor (1*32)(src), tiv, x1; \
144 	vpxor (2*32)(src), tiv, x2; \
148 	vpxor (3*32)(src), tiv, x3; \
152 	vpxor (4*32)(src), tiv, x4; \
156 	vpxor (5*32)(src), tiv, x5; \
160 	vpxor (6*32)(src), tiv, x6; \
164 	vpxor (7*32)(src), tiv, x7; \