40 	pshufd $0, t, t;
42 #define K(x0, x1, x2, x3, x4, i) \  argument
43 	get_key(i, 0, x4); \
50 	pxor x4,		x3;
52 #define LK(x0, x1, x2, x3, x4, i) \  argument
69 	pxor x2,		x3; \
70 	pxor x4,		x3; \
71 	movdqa x3,		x4; \
72 	pslld $7,		x3; \
74 	por x4,			x3; \
78 	pxor x3,		x0; \
79 	pxor x3,		x2; \
85 	pxor RT0,		x3; \
93 	get_key(i, 0, RT0); \
98 #define KL(x0, x1, x2, x3, x4, i) \  argument
99 	K(x0, x1, x2, x3, x4, i); \
108 	pxor x3,		x2; \
109 	pxor x3,		x0; \
118 	movdqa x3,		x4; \
119 	psrld $7,		x3; \
121 	por x4,			x3; \
125 	pxor x4,		x3; \
131 	pxor x2,		x3; \
137 #define S0(x0, x1, x2, x3, x4) \  argument
138 	movdqa x3,		x4; \
139 	por x0,			x3; \
143 	pxor x1,		x3; \
147 	pxor x3,		x0; \
151 	pxor x2,		x3; \
156 #define S1(x0, x1, x2, x3, x4) \  argument
159 	pxor x3,		x0; \
160 	pxor RNOT,		x3; \
163 	pxor x2,		x3; \
164 	pxor x3,		x0; \
165 	pxor x3,		x1; \
166 	pxor x4,		x3; \
176 #define S2(x0, x1, x2, x3, x4) \  argument
177 	pxor RNOT,		x3; \
181 	pxor x3,		x0; \
182 	por x4,			x3; \
184 	pxor x1,		x3; \
187 	pand x3,		x2; \
188 	por x1,			x3; \
190 	pxor x0,		x3; \
195 #define S3(x0, x1, x2, x3, x4) \  argument
197 	pxor x3,		x1; \
198 	por x0,			x3; \
202 	pand x3,		x1; \
203 	pxor x3,		x2; \
205 	pxor x3,		x4; \
207 	pand x3,		x0; \
208 	pand x4,		x3; \
209 	pxor x2,		x3; \
212 	pxor x3,		x4; \
213 	pxor x3,		x0; \
214 	pxor x2,		x3;
216 #define S4(x0, x1, x2, x3, x4) \  argument
217 	movdqa x3,		x4; \
218 	pand x0,		x3; \
220 	pxor x2,		x3; \
223 	pxor x3,		x4; \
229 	pxor x3,		x2; \
231 	por x1,			x3; \
233 	pxor x0,		x3;
235 #define S5(x0, x1, x2, x3, x4) \  argument
239 	pxor RNOT,		x3; \
243 	por x3,			x4; \
245 	pand x3,		x0; \
246 	pxor x3,		x1; \
247 	pxor x2,		x3; \
252 	pxor x2,		x3;
254 #define S6(x0, x1, x2, x3, x4) \  argument
256 	pxor x0,		x3; \
259 	pand x3,		x0; \
260 	por x3,			x1; \
264 	pxor x4,		x3; \
268 	pxor x3,		x2; \
269 	pand x1,		x3; \
270 	pxor x0,		x3; \
273 #define S7(x0, x1, x2, x3, x4) \  argument
278 	pxor x3,		x1; \
279 	por x4,			x3; \
281 	pxor x3,		x2; \
282 	pxor x0,		x3; \
286 	pxor x3,		x4; \
287 	pand x0,		x3; \
290 	pxor x1,		x3; \
294 #define SI0(x0, x1, x2, x3, x4) \  argument
295 	movdqa x3,		x4; \
297 	por x1,			x3; \
300 	pxor x3,		x2; \
301 	pxor x0,		x3; \
304 	pand x3,		x2; \
305 	pxor x4,		x3; \
306 	pxor x3,		x2; \
307 	pxor x3,		x1; \
308 	pand x0,		x3; \
311 	pxor x3,		x4;
313 #define SI1(x0, x1, x2, x3, x4) \  argument
314 	pxor x3,		x1; \
319 	pxor x3,		x4; \
320 	pand x1,		x3; \
324 	por x3,			x1; \
325 	pxor x0,		x3; \
332 #define SI2(x0, x1, x2, x3, x4) \  argument
334 	movdqa x3,		x4; \
335 	pxor RNOT,		x3; \
336 	por x2,			x3; \
339 	pxor x1,		x3; \
343 	por x3,			x4; \
344 	pxor x3,		x2; \
347 	pxor x3,		x2; \
348 	pxor x4,		x3; \
351 #define SI3(x0, x1, x2, x3, x4) \  argument
357 	pxor x3,		x4; \
358 	pxor x3,		x0; \
359 	por x1,			x3; \
361 	pxor x3,		x1; \
363 	pxor x3,		x2; \
364 	pand x1,		x3; \
367 	pxor x3,		x4; \
368 	pxor x0,		x3; \
371 #define SI4(x0, x1, x2, x3, x4) \  argument
372 	pxor x3,		x2; \
376 	por x3,			x2; \
383 	pxor x3,		x0; \
384 	pand x2,		x3; \
385 	pxor x3,		x4; \
386 	pxor x1,		x3; \
389 	pxor x3,		x0;
391 #define SI5(x0, x1, x2, x3, x4) \  argument
395 	pxor x3,		x1; \
396 	pand x4,		x3; \
397 	pxor x3,		x2; \
398 	por x0,			x3; \
400 	pxor x2,		x3; \
406 	pxor x3,		x1; \
408 	pxor x3,		x2; \
411 	pxor x3,		x4;
413 #define SI6(x0, x1, x2, x3, x4) \  argument
416 	pand x3,		x0; \
417 	pxor x3,		x2; \
419 	pxor x1,		x3; \
421 	pxor x3,		x2; \
422 	pand x0,		x3; \
424 	pxor x1,		x3; \
427 	pxor x4,		x3; \
432 #define SI7(x0, x1, x2, x3, x4) \  argument
433 	movdqa x3,		x4; \
434 	pand x0,		x3; \
439 	por x3,			x1; \
444 	pxor x2,		x3; \
445 	pxor x3,		x4; \
446 	pand x3,		x2; \
447 	por x0,			x3; \
449 	pxor x4,		x3; \
453 #define transpose_4x4(x0, x1, x2, x3, t0, t1, t2) \  argument
458 	punpckhdq x3,		x2; \
459 	punpckldq x3,		t1; \
463 	movdqa t2,		x3; \
465 	punpckhqdq x2,		x3; \
468 #define read_blocks(in, x0, x1, x2, x3, t0, t1, t2) \  argument
469 	movdqu (0*4*4)(in),	x0; \
472 	movdqu (3*4*4)(in),	x3; \
474 	transpose_4x4(x0, x1, x2, x3, t0, t1, t2)
476 #define write_blocks(out, x0, x1, x2, x3, t0, t1, t2) \  argument
477 	transpose_4x4(x0, x1, x2, x3, t0, t1, t2) \
479 	movdqu x0, (0*4*4)(out); \
482 	movdqu x3, (3*4*4)(out);
484 #define xor_blocks(out, x0, x1, x2, x3, t0, t1, t2) \  argument
485 	transpose_4x4(x0, x1, x2, x3, t0, t1, t2) \
487 	movdqu (0*4*4)(out),	t0; \
489 	movdqu x0,		(0*4*4)(out); \
497 	pxor t0,		x3; \
498 	movdqu x3,		(3*4*4)(out);
515 					 K(RA, RB, RC, RD, RE, 0);
551 	cmpb $0, arg_xor(%esp);
610 	SI0(RE, RB, RC, RA, RD);	 K(RC, RD, RB, RE, RA, 0);