32 #define loadkeys(x0, x1, x2, x3, i) \  argument
33 	({ x0 = k[i]; x1 = k[i+1]; x2 = k[i+2]; x3 = k[i+3]; })
35 #define storekeys(x0, x1, x2, x3, i) \  argument
36 	({ k[i] = x0; k[i+1] = x1; k[i+2] = x2; k[i+3] = x3; })
38 #define store_and_load_keys(x0, x1, x2, x3, s, l) \  argument
39 	({ storekeys(x0, x1, x2, x3, s); loadkeys(x0, x1, x2, x3, l); })
41 #define K(x0, x1, x2, x3, i) ({				\  argument
42 	x3 ^= k[4*(i)+3];        x2 ^= k[4*(i)+2];	\
46 #define LK(x0, x1, x2, x3, x4, i) ({					   \  argument
49 	x3 ^= x2;		x1 ^= x2;				   \
50 	x1 = rol32(x1, 1);	x3 ^= x4;				   \
51 	x3 = rol32(x3, 7);	x4  = x1;				   \
52 	x0 ^= x1;		x4 <<= 7;		x2 ^= x3;	   \
53 	x0 ^= x3;		x2 ^= x4;		x3 ^= k[4*i+3];	   \
58 #define KL(x0, x1, x2, x3, x4, i) ({					   \  argument
60 	x3 ^= k[4*i+3];		x0 = ror32(x0, 5);	x2 = ror32(x2, 22);\
61 	x4 =  x1;		x2 ^= x3;		x0 ^= x3;	   \
63 	x2 ^= x4;		x3 = ror32(x3, 7);	x4 = x0 << 3;	   \
64 	x1 ^= x0;		x3 ^= x4;		x0 = ror32(x0, 13);\
65 	x1 ^= x2;		x3 ^= x2;		x2 = ror32(x2, 3); \
68 #define S0(x0, x1, x2, x3, x4) ({			\  argument
69 					x4  = x3;	\
70 	x3 |= x0;	x0 ^= x4;	x4 ^= x2;	\
71 	x4 = ~x4;	x3 ^= x1;	x1 &= x0;	\
72 	x1 ^= x4;	x2 ^= x0;	x0 ^= x3;	\
74 	x3 ^= x2;	x1 = ~x1;	x2 ^= x4;	\
78 #define S1(x0, x1, x2, x3, x4) ({			\  argument
80 	x1 ^= x0;	x0 ^= x3;	x3 = ~x3;	\
81 	x4 &= x1;	x0 |= x1;	x3 ^= x2;	\
82 	x0 ^= x3;	x1 ^= x3;	x3 ^= x4;	\
88 #define S2(x0, x1, x2, x3, x4) ({			\  argument
89 					x3 = ~x3;	\
91 	x0 ^= x3;	x3 |= x4;	x2 ^= x1;	\
92 	x3 ^= x1;	x1 &= x0;	x0 ^= x2;	\
93 	x2 &= x3;	x3 |= x1;	x0 = ~x0;	\
94 	x3 ^= x0;	x4 ^= x0;	x0 ^= x2;	\
98 #define S3(x0, x1, x2, x3, x4) ({			\  argument
100 	x1 ^= x3;	x3 |= x0;	x4 &= x0;	\
101 	x0 ^= x2;	x2 ^= x1;	x1 &= x3;	\
102 	x2 ^= x3;	x0 |= x4;	x4 ^= x3;	\
103 	x1 ^= x0;	x0 &= x3;	x3 &= x4;	\
104 	x3 ^= x2;	x4 |= x1;	x2 &= x1;	\
105 	x4 ^= x3;	x0 ^= x3;	x3 ^= x2;	\
108 #define S4(x0, x1, x2, x3, x4) ({			\  argument
109 					x4  = x3;	\
110 	x3 &= x0;	x0 ^= x4;			\
111 	x3 ^= x2;	x2 |= x4;	x0 ^= x1;	\
112 	x4 ^= x3;	x2 |= x0;			\
114 	x1 ^= x4;	x4 &= x2;	x2 ^= x3;	\
115 	x4 ^= x0;	x3 |= x1;	x1 = ~x1;	\
116 	x3 ^= x0;					\
119 #define S5(x0, x1, x2, x3, x4) ({			\  argument
121 	x2 ^= x1;	x3 = ~x3;	x4 ^= x0;	\
122 	x0 ^= x2;	x1 &= x4;	x4 |= x3;	\
123 	x4 ^= x0;	x0 &= x3;	x1 ^= x3;	\
124 	x3 ^= x2;	x0 ^= x1;	x2 &= x4;	\
126 	x3 ^= x2;					\
129 #define S6(x0, x1, x2, x3, x4) ({			\  argument
131 	x3 ^= x0;	x1 ^= x2;	x2 ^= x0;	\
132 	x0 &= x3;	x1 |= x3;	x4 = ~x4;	\
134 	x3 ^= x4;	x4 ^= x0;	x2 &= x0;	\
135 	x4 ^= x1;	x2 ^= x3;	x3 &= x1;	\
136 	x3 ^= x0;	x1 ^= x2;			\
139 #define S7(x0, x1, x2, x3, x4) ({			\  argument
142 	x1 ^= x3;	x3 |= x4;	x4 ^= x2;	\
143 	x2 ^= x3;	x3 ^= x0;	x0 |= x1;	\
144 	x2 &= x0;	x0 ^= x4;	x4 ^= x3;	\
145 	x3 &= x0;	x4 ^= x1;			\
146 	x2 ^= x4;	x3 ^= x1;	x4 |= x0;	\
150 #define SI0(x0, x1, x2, x3, x4) ({			\  argument
151 			x4  = x3;	x1 ^= x0;	\
152 	x3 |= x1;	x4 ^= x1;	x0 = ~x0;	\
153 	x2 ^= x3;	x3 ^= x0;	x0 &= x1;	\
154 	x0 ^= x2;	x2 &= x3;	x3 ^= x4;	\
155 	x2 ^= x3;	x1 ^= x3;	x3 &= x0;	\
156 	x1 ^= x0;	x0 ^= x2;	x4 ^= x3;	\
159 #define SI1(x0, x1, x2, x3, x4) ({			\  argument
160 	x1 ^= x3;	x4  = x0;			\
162 	x4 ^= x3;	x3 &= x1;	x1 ^= x2;	\
163 	x2 &= x4;	x4 ^= x1;	x1 |= x3;	\
164 	x3 ^= x0;	x2 ^= x0;	x0 |= x4;	\
169 #define SI2(x0, x1, x2, x3, x4) ({			\  argument
170 	x2 ^= x1;	x4  = x3;	x3 = ~x3;	\
171 	x3 |= x2;	x2 ^= x4;	x4 ^= x0;	\
172 	x3 ^= x1;	x1 |= x2;	x2 ^= x0;	\
173 	x1 ^= x4;	x4 |= x3;	x2 ^= x3;	\
175 	x2 ^= x3;	x3 ^= x4;	x4 ^= x0;	\
178 #define SI3(x0, x1, x2, x3, x4) ({			\  argument
181 	x1 ^= x0;	x0 |= x4;	x4 ^= x3;	\
182 	x0 ^= x3;	x3 |= x1;	x1 ^= x2;	\
183 	x1 ^= x3;	x0 ^= x2;	x2 ^= x3;	\
184 	x3 &= x1;	x1 ^= x0;	x0 &= x2;	\
185 	x4 ^= x3;	x3 ^= x0;	x0 ^= x1;	\
188 #define SI4(x0, x1, x2, x3, x4) ({			\  argument
189 	x2 ^= x3;	x4  = x0;	x0 &= x1;	\
190 	x0 ^= x2;	x2 |= x3;	x4 = ~x4;	\
193 	x0 ^= x3;	x3 &= x2;			\
194 	x4 ^= x3;	x3 ^= x1;	x1 &= x0;	\
195 	x4 ^= x1;	x0 ^= x3;			\
198 #define SI5(x0, x1, x2, x3, x4) ({			\  argument
200 	x2 ^= x4;	x1 ^= x3;	x3 &= x4;	\
201 	x2 ^= x3;	x3 |= x0;	x0 = ~x0;	\
202 	x3 ^= x2;	x2 |= x0;	x4 ^= x1;	\
204 	x1 ^= x3;	x0 &= x2;	x2 ^= x3;	\
205 	x0 ^= x2;	x2 ^= x4;	x4 ^= x3;	\
208 #define SI6(x0, x1, x2, x3, x4) ({			\  argument
210 	x4  = x0;	x0 &= x3;	x2 ^= x3;	\
211 	x0 ^= x2;	x3 ^= x1;	x2 |= x4;	\
212 	x2 ^= x3;	x3 &= x0;	x0 = ~x0;	\
213 	x3 ^= x1;	x1 &= x2;	x4 ^= x0;	\
214 	x3 ^= x4;	x4 ^= x2;	x0 ^= x1;	\
218 #define SI7(x0, x1, x2, x3, x4) ({			\  argument
219 	x4  = x3;	x3 &= x0;	x0 ^= x2;	\
221 	x1 |= x3;	x4 ^= x0;	x0 &= x2;	\
222 	x0 ^= x1;	x1 &= x2;	x3 ^= x2;	\
223 	x4 ^= x3;	x2 &= x3;	x3 |= x0;	\
224 	x1 ^= x4;	x3 ^= x4;	x4 &= x0;	\