16 #define out_blk   8  /* output byte array address parameter*/
21 #define c_offset	8
30 #define w	4096	/* 8 whitening keys (word) */
33 /* define a few register aliases to allow macro substitution */
61  * a input register containing a (rotated 16)
65  * operations on a and b are interleaved to increase performance
67 #define encrypt_round(a,b,c,d,round)\  argument
71 	movzx	a ## B,		%edi;\
76 	movzx	a ## H,		%edi;\
77 	ror	$16,		a ## D;\
81 	movzx	a ## B,		%edi;\
86 	movzx	a ## H,		%edi;\
98  * a input register containing a (rotated 16)
102  * operations on a and b are interleaved to increase performance
105 #define encrypt_last_round(a,b,c,d,round)\  argument
109 	movzx	a ## B,		%edi;\
114 	movzx	a ## H,		%edi;\
115 	ror	$16,		a ## D;\
119 	movzx	a ## B,		%edi;\
124 	movzx	a ## H,		%edi;\
136  * a input register containing a
140  * operations on a and b are interleaved to increase performance
142 #define decrypt_round(a,b,c,d,round)\  argument
144 	movzx	a ## B,		%edi;\
148 	movzx	a ## H,		%edi;\
149 	ror	$16,		a ## D;\
154 	movzx	a ## B,		%edi;\
158 	movzx	a ## H,		%edi;\
159 	ror	$15,		a ## D;\
173  * a input register containing a
177  * operations on a and b are interleaved to increase performance
180 #define decrypt_last_round(a,b,c,d,round)\  argument
182 	movzx	a ## B,		%edi;\
186 	movzx	a ## H,		%edi;\
187 	ror	$16,		a ## D;\
192 	movzx	a ## B,		%edi;\
196 	movzx	a ## H,		%edi;\
197 	ror	$16,		a ## D;\
232 	encrypt_round(R2,R3,R0,R1,8);
233 	encrypt_round(R0,R1,R2,R3,2*8);
234 	encrypt_round(R2,R3,R0,R1,3*8);
235 	encrypt_round(R0,R1,R2,R3,4*8);
236 	encrypt_round(R2,R3,R0,R1,5*8);
237 	encrypt_round(R0,R1,R2,R3,6*8);
238 	encrypt_round(R2,R3,R0,R1,7*8);
239 	encrypt_round(R0,R1,R2,R3,8*8);
240 	encrypt_round(R2,R3,R0,R1,9*8);
241 	encrypt_round(R0,R1,R2,R3,10*8);
242 	encrypt_round(R2,R3,R0,R1,11*8);
243 	encrypt_round(R0,R1,R2,R3,12*8);
244 	encrypt_round(R2,R3,R0,R1,13*8);
245 	encrypt_round(R0,R1,R2,R3,14*8);
246 	encrypt_last_round(R2,R3,R0,R1,15*8);
288 	decrypt_round(R0,R1,R2,R3,15*8);
289 	decrypt_round(R2,R3,R0,R1,14*8);
290 	decrypt_round(R0,R1,R2,R3,13*8);
291 	decrypt_round(R2,R3,R0,R1,12*8);
292 	decrypt_round(R0,R1,R2,R3,11*8);
293 	decrypt_round(R2,R3,R0,R1,10*8);
294 	decrypt_round(R0,R1,R2,R3,9*8);
295 	decrypt_round(R2,R3,R0,R1,8*8);
296 	decrypt_round(R0,R1,R2,R3,7*8);
297 	decrypt_round(R2,R3,R0,R1,6*8);
298 	decrypt_round(R0,R1,R2,R3,5*8);
299 	decrypt_round(R2,R3,R0,R1,4*8);
300 	decrypt_round(R0,R1,R2,R3,3*8);
301 	decrypt_round(R2,R3,R0,R1,2*8);
302 	decrypt_round(R0,R1,R2,R3,1*8);