8 #define load_8way(src, x0, x1, x2, x3, x4, x5, x6, x7) \  argument
9 	vmovdqu (0*16)(src), x0; \
16 	vmovdqu (7*16)(src), x7;
18 #define store_8way(dst, x0, x1, x2, x3, x4, x5, x6, x7) \  argument
19 	vmovdqu x0, (0*16)(dst); \
26 	vmovdqu x7, (7*16)(dst);
28 #define store_cbc_8way(src, dst, x0, x1, x2, x3, x4, x5, x6, x7) \  argument
29 	vpxor (0*16)(src), x1, x1; \
35 	vpxor (6*16)(src), x7, x7; \
36 	store_8way(dst, x0, x1, x2, x3, x4, x5, x6, x7);
44 #define load_ctr_8way(iv, bswap, x0, x1, x2, x3, x4, x5, x6, x7, t0, t1, t2) \  argument
46 	vpsrldq $8, t0, t0; /* low: -1, high: 0 */ \
50 	vmovdqu (iv), x7; \
51 	vpshufb t1, x7, x0; \
54 	inc_le128(x7, t0, t2); \
55 	vpshufb t1, x7, x1; \
56 	inc_le128(x7, t0, t2); \
57 	vpshufb t1, x7, x2; \
58 	inc_le128(x7, t0, t2); \
59 	vpshufb t1, x7, x3; \
60 	inc_le128(x7, t0, t2); \
61 	vpshufb t1, x7, x4; \
62 	inc_le128(x7, t0, t2); \
63 	vpshufb t1, x7, x5; \
64 	inc_le128(x7, t0, t2); \
65 	vpshufb t1, x7, x6; \
66 	inc_le128(x7, t0, t2); \
67 	vmovdqa x7, t2; \
68 	vpshufb t1, 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; \
80 	vpxor (7*16)(src), x7, x7; \
81 	store_8way(dst, x0, x1, x2, x3, x4, x5, x6, x7);
86 	vpshufd $0x13, tmp, tmp; \
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; \
97 	vmovdqu tiv, (0*16)(dst); \
125 	vpxor (7*16)(src), tiv, x7; \
131 #define store_xts_8way(dst, x0, x1, x2, x3, x4, x5, x6, x7) \  argument
132 	vpxor (0*16)(dst), x0, x0; \
139 	vpxor (7*16)(dst), x7, x7; \
140 	store_8way(dst, x0, x1, x2, x3, x4, x5, x6, x7);