aesni-x86.S@ 97371

Last change on this file since 97371 was 94083, checked in by vboxsync, 3 years ago
libs/openssl-3.0.1: Recreate asm files, bugref:10128
File size: 62.9 KB

Line
1	.text
2	.globl _aesni_encrypt
3	.align 4
4	_aesni_encrypt:
5	L_aesni_encrypt_begin:
6	.byte 243,15,30,251
7	movl 4(%esp),%eax
8	movl 12(%esp),%edx
9	movups (%eax),%xmm2
10	movl 240(%edx),%ecx
11	movl 8(%esp),%eax
12	movups (%edx),%xmm0
13	movups 16(%edx),%xmm1
14	leal 32(%edx),%edx
15	xorps %xmm0,%xmm2
16	L000enc1_loop_1:
17	.byte 102,15,56,220,209
18	decl %ecx
19	movups (%edx),%xmm1
20	leal 16(%edx),%edx
21	jnz L000enc1_loop_1
22	.byte 102,15,56,221,209
23	pxor %xmm0,%xmm0
24	pxor %xmm1,%xmm1
25	movups %xmm2,(%eax)
26	pxor %xmm2,%xmm2
27	ret
28	.globl _aesni_decrypt
29	.align 4
30	_aesni_decrypt:
31	L_aesni_decrypt_begin:
32	.byte 243,15,30,251
33	movl 4(%esp),%eax
34	movl 12(%esp),%edx
35	movups (%eax),%xmm2
36	movl 240(%edx),%ecx
37	movl 8(%esp),%eax
38	movups (%edx),%xmm0
39	movups 16(%edx),%xmm1
40	leal 32(%edx),%edx
41	xorps %xmm0,%xmm2
42	L001dec1_loop_2:
43	.byte 102,15,56,222,209
44	decl %ecx
45	movups (%edx),%xmm1
46	leal 16(%edx),%edx
47	jnz L001dec1_loop_2
48	.byte 102,15,56,223,209
49	pxor %xmm0,%xmm0
50	pxor %xmm1,%xmm1
51	movups %xmm2,(%eax)
52	pxor %xmm2,%xmm2
53	ret
54	.align 4
55	__aesni_encrypt2:
56	.byte 243,15,30,251
57	movups (%edx),%xmm0
58	shll $4,%ecx
59	movups 16(%edx),%xmm1
60	xorps %xmm0,%xmm2
61	pxor %xmm0,%xmm3
62	movups 32(%edx),%xmm0
63	leal 32(%edx,%ecx,1),%edx
64	negl %ecx
65	addl $16,%ecx
66	L002enc2_loop:
67	.byte 102,15,56,220,209
68	.byte 102,15,56,220,217
69	movups (%edx,%ecx,1),%xmm1
70	addl $32,%ecx
71	.byte 102,15,56,220,208
72	.byte 102,15,56,220,216
73	movups -16(%edx,%ecx,1),%xmm0
74	jnz L002enc2_loop
75	.byte 102,15,56,220,209
76	.byte 102,15,56,220,217
77	.byte 102,15,56,221,208
78	.byte 102,15,56,221,216
79	ret
80	.align 4
81	__aesni_decrypt2:
82	.byte 243,15,30,251
83	movups (%edx),%xmm0
84	shll $4,%ecx
85	movups 16(%edx),%xmm1
86	xorps %xmm0,%xmm2
87	pxor %xmm0,%xmm3
88	movups 32(%edx),%xmm0
89	leal 32(%edx,%ecx,1),%edx
90	negl %ecx
91	addl $16,%ecx
92	L003dec2_loop:
93	.byte 102,15,56,222,209
94	.byte 102,15,56,222,217
95	movups (%edx,%ecx,1),%xmm1
96	addl $32,%ecx
97	.byte 102,15,56,222,208
98	.byte 102,15,56,222,216
99	movups -16(%edx,%ecx,1),%xmm0
100	jnz L003dec2_loop
101	.byte 102,15,56,222,209
102	.byte 102,15,56,222,217
103	.byte 102,15,56,223,208
104	.byte 102,15,56,223,216
105	ret
106	.align 4
107	__aesni_encrypt3:
108	.byte 243,15,30,251
109	movups (%edx),%xmm0
110	shll $4,%ecx
111	movups 16(%edx),%xmm1
112	xorps %xmm0,%xmm2
113	pxor %xmm0,%xmm3
114	pxor %xmm0,%xmm4
115	movups 32(%edx),%xmm0
116	leal 32(%edx,%ecx,1),%edx
117	negl %ecx
118	addl $16,%ecx
119	L004enc3_loop:
120	.byte 102,15,56,220,209
121	.byte 102,15,56,220,217
122	.byte 102,15,56,220,225
123	movups (%edx,%ecx,1),%xmm1
124	addl $32,%ecx
125	.byte 102,15,56,220,208
126	.byte 102,15,56,220,216
127	.byte 102,15,56,220,224
128	movups -16(%edx,%ecx,1),%xmm0
129	jnz L004enc3_loop
130	.byte 102,15,56,220,209
131	.byte 102,15,56,220,217
132	.byte 102,15,56,220,225
133	.byte 102,15,56,221,208
134	.byte 102,15,56,221,216
135	.byte 102,15,56,221,224
136	ret
137	.align 4
138	__aesni_decrypt3:
139	.byte 243,15,30,251
140	movups (%edx),%xmm0
141	shll $4,%ecx
142	movups 16(%edx),%xmm1
143	xorps %xmm0,%xmm2
144	pxor %xmm0,%xmm3
145	pxor %xmm0,%xmm4
146	movups 32(%edx),%xmm0
147	leal 32(%edx,%ecx,1),%edx
148	negl %ecx
149	addl $16,%ecx
150	L005dec3_loop:
151	.byte 102,15,56,222,209
152	.byte 102,15,56,222,217
153	.byte 102,15,56,222,225
154	movups (%edx,%ecx,1),%xmm1
155	addl $32,%ecx
156	.byte 102,15,56,222,208
157	.byte 102,15,56,222,216
158	.byte 102,15,56,222,224
159	movups -16(%edx,%ecx,1),%xmm0
160	jnz L005dec3_loop
161	.byte 102,15,56,222,209
162	.byte 102,15,56,222,217
163	.byte 102,15,56,222,225
164	.byte 102,15,56,223,208
165	.byte 102,15,56,223,216
166	.byte 102,15,56,223,224
167	ret
168	.align 4
169	__aesni_encrypt4:
170	.byte 243,15,30,251
171	movups (%edx),%xmm0
172	movups 16(%edx),%xmm1
173	shll $4,%ecx
174	xorps %xmm0,%xmm2
175	pxor %xmm0,%xmm3
176	pxor %xmm0,%xmm4
177	pxor %xmm0,%xmm5
178	movups 32(%edx),%xmm0
179	leal 32(%edx,%ecx,1),%edx
180	negl %ecx
181	.byte 15,31,64,0
182	addl $16,%ecx
183	L006enc4_loop:
184	.byte 102,15,56,220,209
185	.byte 102,15,56,220,217
186	.byte 102,15,56,220,225
187	.byte 102,15,56,220,233
188	movups (%edx,%ecx,1),%xmm1
189	addl $32,%ecx
190	.byte 102,15,56,220,208
191	.byte 102,15,56,220,216
192	.byte 102,15,56,220,224
193	.byte 102,15,56,220,232
194	movups -16(%edx,%ecx,1),%xmm0
195	jnz L006enc4_loop
196	.byte 102,15,56,220,209
197	.byte 102,15,56,220,217
198	.byte 102,15,56,220,225
199	.byte 102,15,56,220,233
200	.byte 102,15,56,221,208
201	.byte 102,15,56,221,216
202	.byte 102,15,56,221,224
203	.byte 102,15,56,221,232
204	ret
205	.align 4
206	__aesni_decrypt4:
207	.byte 243,15,30,251
208	movups (%edx),%xmm0
209	movups 16(%edx),%xmm1
210	shll $4,%ecx
211	xorps %xmm0,%xmm2
212	pxor %xmm0,%xmm3
213	pxor %xmm0,%xmm4
214	pxor %xmm0,%xmm5
215	movups 32(%edx),%xmm0
216	leal 32(%edx,%ecx,1),%edx
217	negl %ecx
218	.byte 15,31,64,0
219	addl $16,%ecx
220	L007dec4_loop:
221	.byte 102,15,56,222,209
222	.byte 102,15,56,222,217
223	.byte 102,15,56,222,225
224	.byte 102,15,56,222,233
225	movups (%edx,%ecx,1),%xmm1
226	addl $32,%ecx
227	.byte 102,15,56,222,208
228	.byte 102,15,56,222,216
229	.byte 102,15,56,222,224
230	.byte 102,15,56,222,232
231	movups -16(%edx,%ecx,1),%xmm0
232	jnz L007dec4_loop
233	.byte 102,15,56,222,209
234	.byte 102,15,56,222,217
235	.byte 102,15,56,222,225
236	.byte 102,15,56,222,233
237	.byte 102,15,56,223,208
238	.byte 102,15,56,223,216
239	.byte 102,15,56,223,224
240	.byte 102,15,56,223,232
241	ret
242	.align 4
243	__aesni_encrypt6:
244	.byte 243,15,30,251
245	movups (%edx),%xmm0
246	shll $4,%ecx
247	movups 16(%edx),%xmm1
248	xorps %xmm0,%xmm2
249	pxor %xmm0,%xmm3
250	pxor %xmm0,%xmm4
251	.byte 102,15,56,220,209
252	pxor %xmm0,%xmm5
253	pxor %xmm0,%xmm6
254	.byte 102,15,56,220,217
255	leal 32(%edx,%ecx,1),%edx
256	negl %ecx
257	.byte 102,15,56,220,225
258	pxor %xmm0,%xmm7
259	movups (%edx,%ecx,1),%xmm0
260	addl $16,%ecx
261	jmp L008_aesni_encrypt6_inner
262	.align 4,0x90
263	L009enc6_loop:
264	.byte 102,15,56,220,209
265	.byte 102,15,56,220,217
266	.byte 102,15,56,220,225
267	L008_aesni_encrypt6_inner:
268	.byte 102,15,56,220,233
269	.byte 102,15,56,220,241
270	.byte 102,15,56,220,249
271	L_aesni_encrypt6_enter:
272	movups (%edx,%ecx,1),%xmm1
273	addl $32,%ecx
274	.byte 102,15,56,220,208
275	.byte 102,15,56,220,216
276	.byte 102,15,56,220,224
277	.byte 102,15,56,220,232
278	.byte 102,15,56,220,240
279	.byte 102,15,56,220,248
280	movups -16(%edx,%ecx,1),%xmm0
281	jnz L009enc6_loop
282	.byte 102,15,56,220,209
283	.byte 102,15,56,220,217
284	.byte 102,15,56,220,225
285	.byte 102,15,56,220,233
286	.byte 102,15,56,220,241
287	.byte 102,15,56,220,249
288	.byte 102,15,56,221,208
289	.byte 102,15,56,221,216
290	.byte 102,15,56,221,224
291	.byte 102,15,56,221,232
292	.byte 102,15,56,221,240
293	.byte 102,15,56,221,248
294	ret
295	.align 4
296	__aesni_decrypt6:
297	.byte 243,15,30,251
298	movups (%edx),%xmm0
299	shll $4,%ecx
300	movups 16(%edx),%xmm1
301	xorps %xmm0,%xmm2
302	pxor %xmm0,%xmm3
303	pxor %xmm0,%xmm4
304	.byte 102,15,56,222,209
305	pxor %xmm0,%xmm5
306	pxor %xmm0,%xmm6
307	.byte 102,15,56,222,217
308	leal 32(%edx,%ecx,1),%edx
309	negl %ecx
310	.byte 102,15,56,222,225
311	pxor %xmm0,%xmm7
312	movups (%edx,%ecx,1),%xmm0
313	addl $16,%ecx
314	jmp L010_aesni_decrypt6_inner
315	.align 4,0x90
316	L011dec6_loop:
317	.byte 102,15,56,222,209
318	.byte 102,15,56,222,217
319	.byte 102,15,56,222,225
320	L010_aesni_decrypt6_inner:
321	.byte 102,15,56,222,233
322	.byte 102,15,56,222,241
323	.byte 102,15,56,222,249
324	L_aesni_decrypt6_enter:
325	movups (%edx,%ecx,1),%xmm1
326	addl $32,%ecx
327	.byte 102,15,56,222,208
328	.byte 102,15,56,222,216
329	.byte 102,15,56,222,224
330	.byte 102,15,56,222,232
331	.byte 102,15,56,222,240
332	.byte 102,15,56,222,248
333	movups -16(%edx,%ecx,1),%xmm0
334	jnz L011dec6_loop
335	.byte 102,15,56,222,209
336	.byte 102,15,56,222,217
337	.byte 102,15,56,222,225
338	.byte 102,15,56,222,233
339	.byte 102,15,56,222,241
340	.byte 102,15,56,222,249
341	.byte 102,15,56,223,208
342	.byte 102,15,56,223,216
343	.byte 102,15,56,223,224
344	.byte 102,15,56,223,232
345	.byte 102,15,56,223,240
346	.byte 102,15,56,223,248
347	ret
348	.globl _aesni_ecb_encrypt
349	.align 4
350	_aesni_ecb_encrypt:
351	L_aesni_ecb_encrypt_begin:
352	.byte 243,15,30,251
353	pushl %ebp
354	pushl %ebx
355	pushl %esi
356	pushl %edi
357	movl 20(%esp),%esi
358	movl 24(%esp),%edi
359	movl 28(%esp),%eax
360	movl 32(%esp),%edx
361	movl 36(%esp),%ebx
362	andl $-16,%eax
363	jz L012ecb_ret
364	movl 240(%edx),%ecx
365	testl %ebx,%ebx
366	jz L013ecb_decrypt
367	movl %edx,%ebp
368	movl %ecx,%ebx
369	cmpl $96,%eax
370	jb L014ecb_enc_tail
371	movdqu (%esi),%xmm2
372	movdqu 16(%esi),%xmm3
373	movdqu 32(%esi),%xmm4
374	movdqu 48(%esi),%xmm5
375	movdqu 64(%esi),%xmm6
376	movdqu 80(%esi),%xmm7
377	leal 96(%esi),%esi
378	subl $96,%eax
379	jmp L015ecb_enc_loop6_enter
380	.align 4,0x90
381	L016ecb_enc_loop6:
382	movups %xmm2,(%edi)
383	movdqu (%esi),%xmm2
384	movups %xmm3,16(%edi)
385	movdqu 16(%esi),%xmm3
386	movups %xmm4,32(%edi)
387	movdqu 32(%esi),%xmm4
388	movups %xmm5,48(%edi)
389	movdqu 48(%esi),%xmm5
390	movups %xmm6,64(%edi)
391	movdqu 64(%esi),%xmm6
392	movups %xmm7,80(%edi)
393	leal 96(%edi),%edi
394	movdqu 80(%esi),%xmm7
395	leal 96(%esi),%esi
396	L015ecb_enc_loop6_enter:
397	call __aesni_encrypt6
398	movl %ebp,%edx
399	movl %ebx,%ecx
400	subl $96,%eax
401	jnc L016ecb_enc_loop6
402	movups %xmm2,(%edi)
403	movups %xmm3,16(%edi)
404	movups %xmm4,32(%edi)
405	movups %xmm5,48(%edi)
406	movups %xmm6,64(%edi)
407	movups %xmm7,80(%edi)
408	leal 96(%edi),%edi
409	addl $96,%eax
410	jz L012ecb_ret
411	L014ecb_enc_tail:
412	movups (%esi),%xmm2
413	cmpl $32,%eax
414	jb L017ecb_enc_one
415	movups 16(%esi),%xmm3
416	je L018ecb_enc_two
417	movups 32(%esi),%xmm4
418	cmpl $64,%eax
419	jb L019ecb_enc_three
420	movups 48(%esi),%xmm5
421	je L020ecb_enc_four
422	movups 64(%esi),%xmm6
423	xorps %xmm7,%xmm7
424	call __aesni_encrypt6
425	movups %xmm2,(%edi)
426	movups %xmm3,16(%edi)
427	movups %xmm4,32(%edi)
428	movups %xmm5,48(%edi)
429	movups %xmm6,64(%edi)
430	jmp L012ecb_ret
431	.align 4,0x90
432	L017ecb_enc_one:
433	movups (%edx),%xmm0
434	movups 16(%edx),%xmm1
435	leal 32(%edx),%edx
436	xorps %xmm0,%xmm2
437	L021enc1_loop_3:
438	.byte 102,15,56,220,209
439	decl %ecx
440	movups (%edx),%xmm1
441	leal 16(%edx),%edx
442	jnz L021enc1_loop_3
443	.byte 102,15,56,221,209
444	movups %xmm2,(%edi)
445	jmp L012ecb_ret
446	.align 4,0x90
447	L018ecb_enc_two:
448	call __aesni_encrypt2
449	movups %xmm2,(%edi)
450	movups %xmm3,16(%edi)
451	jmp L012ecb_ret
452	.align 4,0x90
453	L019ecb_enc_three:
454	call __aesni_encrypt3
455	movups %xmm2,(%edi)
456	movups %xmm3,16(%edi)
457	movups %xmm4,32(%edi)
458	jmp L012ecb_ret
459	.align 4,0x90
460	L020ecb_enc_four:
461	call __aesni_encrypt4
462	movups %xmm2,(%edi)
463	movups %xmm3,16(%edi)
464	movups %xmm4,32(%edi)
465	movups %xmm5,48(%edi)
466	jmp L012ecb_ret
467	.align 4,0x90
468	L013ecb_decrypt:
469	movl %edx,%ebp
470	movl %ecx,%ebx
471	cmpl $96,%eax
472	jb L022ecb_dec_tail
473	movdqu (%esi),%xmm2
474	movdqu 16(%esi),%xmm3
475	movdqu 32(%esi),%xmm4
476	movdqu 48(%esi),%xmm5
477	movdqu 64(%esi),%xmm6
478	movdqu 80(%esi),%xmm7
479	leal 96(%esi),%esi
480	subl $96,%eax
481	jmp L023ecb_dec_loop6_enter
482	.align 4,0x90
483	L024ecb_dec_loop6:
484	movups %xmm2,(%edi)
485	movdqu (%esi),%xmm2
486	movups %xmm3,16(%edi)
487	movdqu 16(%esi),%xmm3
488	movups %xmm4,32(%edi)
489	movdqu 32(%esi),%xmm4
490	movups %xmm5,48(%edi)
491	movdqu 48(%esi),%xmm5
492	movups %xmm6,64(%edi)
493	movdqu 64(%esi),%xmm6
494	movups %xmm7,80(%edi)
495	leal 96(%edi),%edi
496	movdqu 80(%esi),%xmm7
497	leal 96(%esi),%esi
498	L023ecb_dec_loop6_enter:
499	call __aesni_decrypt6
500	movl %ebp,%edx
501	movl %ebx,%ecx
502	subl $96,%eax
503	jnc L024ecb_dec_loop6
504	movups %xmm2,(%edi)
505	movups %xmm3,16(%edi)
506	movups %xmm4,32(%edi)
507	movups %xmm5,48(%edi)
508	movups %xmm6,64(%edi)
509	movups %xmm7,80(%edi)
510	leal 96(%edi),%edi
511	addl $96,%eax
512	jz L012ecb_ret
513	L022ecb_dec_tail:
514	movups (%esi),%xmm2
515	cmpl $32,%eax
516	jb L025ecb_dec_one
517	movups 16(%esi),%xmm3
518	je L026ecb_dec_two
519	movups 32(%esi),%xmm4
520	cmpl $64,%eax
521	jb L027ecb_dec_three
522	movups 48(%esi),%xmm5
523	je L028ecb_dec_four
524	movups 64(%esi),%xmm6
525	xorps %xmm7,%xmm7
526	call __aesni_decrypt6
527	movups %xmm2,(%edi)
528	movups %xmm3,16(%edi)
529	movups %xmm4,32(%edi)
530	movups %xmm5,48(%edi)
531	movups %xmm6,64(%edi)
532	jmp L012ecb_ret
533	.align 4,0x90
534	L025ecb_dec_one:
535	movups (%edx),%xmm0
536	movups 16(%edx),%xmm1
537	leal 32(%edx),%edx
538	xorps %xmm0,%xmm2
539	L029dec1_loop_4:
540	.byte 102,15,56,222,209
541	decl %ecx
542	movups (%edx),%xmm1
543	leal 16(%edx),%edx
544	jnz L029dec1_loop_4
545	.byte 102,15,56,223,209
546	movups %xmm2,(%edi)
547	jmp L012ecb_ret
548	.align 4,0x90
549	L026ecb_dec_two:
550	call __aesni_decrypt2
551	movups %xmm2,(%edi)
552	movups %xmm3,16(%edi)
553	jmp L012ecb_ret
554	.align 4,0x90
555	L027ecb_dec_three:
556	call __aesni_decrypt3
557	movups %xmm2,(%edi)
558	movups %xmm3,16(%edi)
559	movups %xmm4,32(%edi)
560	jmp L012ecb_ret
561	.align 4,0x90
562	L028ecb_dec_four:
563	call __aesni_decrypt4
564	movups %xmm2,(%edi)
565	movups %xmm3,16(%edi)
566	movups %xmm4,32(%edi)
567	movups %xmm5,48(%edi)
568	L012ecb_ret:
569	pxor %xmm0,%xmm0
570	pxor %xmm1,%xmm1
571	pxor %xmm2,%xmm2
572	pxor %xmm3,%xmm3
573	pxor %xmm4,%xmm4
574	pxor %xmm5,%xmm5
575	pxor %xmm6,%xmm6
576	pxor %xmm7,%xmm7
577	popl %edi
578	popl %esi
579	popl %ebx
580	popl %ebp
581	ret
582	.globl _aesni_ccm64_encrypt_blocks
583	.align 4
584	_aesni_ccm64_encrypt_blocks:
585	L_aesni_ccm64_encrypt_blocks_begin:
586	.byte 243,15,30,251
587	pushl %ebp
588	pushl %ebx
589	pushl %esi
590	pushl %edi
591	movl 20(%esp),%esi
592	movl 24(%esp),%edi
593	movl 28(%esp),%eax
594	movl 32(%esp),%edx
595	movl 36(%esp),%ebx
596	movl 40(%esp),%ecx
597	movl %esp,%ebp
598	subl $60,%esp
599	andl $-16,%esp
600	movl %ebp,48(%esp)
601	movdqu (%ebx),%xmm7
602	movdqu (%ecx),%xmm3
603	movl 240(%edx),%ecx
604	movl $202182159,(%esp)
605	movl $134810123,4(%esp)
606	movl $67438087,8(%esp)
607	movl $66051,12(%esp)
608	movl $1,%ebx
609	xorl %ebp,%ebp
610	movl %ebx,16(%esp)
611	movl %ebp,20(%esp)
612	movl %ebp,24(%esp)
613	movl %ebp,28(%esp)
614	shll $4,%ecx
615	movl $16,%ebx
616	leal (%edx),%ebp
617	movdqa (%esp),%xmm5
618	movdqa %xmm7,%xmm2
619	leal 32(%edx,%ecx,1),%edx
620	subl %ecx,%ebx
621	.byte 102,15,56,0,253
622	L030ccm64_enc_outer:
623	movups (%ebp),%xmm0
624	movl %ebx,%ecx
625	movups (%esi),%xmm6
626	xorps %xmm0,%xmm2
627	movups 16(%ebp),%xmm1
628	xorps %xmm6,%xmm0
629	xorps %xmm0,%xmm3
630	movups 32(%ebp),%xmm0
631	L031ccm64_enc2_loop:
632	.byte 102,15,56,220,209
633	.byte 102,15,56,220,217
634	movups (%edx,%ecx,1),%xmm1
635	addl $32,%ecx
636	.byte 102,15,56,220,208
637	.byte 102,15,56,220,216
638	movups -16(%edx,%ecx,1),%xmm0
639	jnz L031ccm64_enc2_loop
640	.byte 102,15,56,220,209
641	.byte 102,15,56,220,217
642	paddq 16(%esp),%xmm7
643	decl %eax
644	.byte 102,15,56,221,208
645	.byte 102,15,56,221,216
646	leal 16(%esi),%esi
647	xorps %xmm2,%xmm6
648	movdqa %xmm7,%xmm2
649	movups %xmm6,(%edi)
650	.byte 102,15,56,0,213
651	leal 16(%edi),%edi
652	jnz L030ccm64_enc_outer
653	movl 48(%esp),%esp
654	movl 40(%esp),%edi
655	movups %xmm3,(%edi)
656	pxor %xmm0,%xmm0
657	pxor %xmm1,%xmm1
658	pxor %xmm2,%xmm2
659	pxor %xmm3,%xmm3
660	pxor %xmm4,%xmm4
661	pxor %xmm5,%xmm5
662	pxor %xmm6,%xmm6
663	pxor %xmm7,%xmm7
664	popl %edi
665	popl %esi
666	popl %ebx
667	popl %ebp
668	ret
669	.globl _aesni_ccm64_decrypt_blocks
670	.align 4
671	_aesni_ccm64_decrypt_blocks:
672	L_aesni_ccm64_decrypt_blocks_begin:
673	.byte 243,15,30,251
674	pushl %ebp
675	pushl %ebx
676	pushl %esi
677	pushl %edi
678	movl 20(%esp),%esi
679	movl 24(%esp),%edi
680	movl 28(%esp),%eax
681	movl 32(%esp),%edx
682	movl 36(%esp),%ebx
683	movl 40(%esp),%ecx
684	movl %esp,%ebp
685	subl $60,%esp
686	andl $-16,%esp
687	movl %ebp,48(%esp)
688	movdqu (%ebx),%xmm7
689	movdqu (%ecx),%xmm3
690	movl 240(%edx),%ecx
691	movl $202182159,(%esp)
692	movl $134810123,4(%esp)
693	movl $67438087,8(%esp)
694	movl $66051,12(%esp)
695	movl $1,%ebx
696	xorl %ebp,%ebp
697	movl %ebx,16(%esp)
698	movl %ebp,20(%esp)
699	movl %ebp,24(%esp)
700	movl %ebp,28(%esp)
701	movdqa (%esp),%xmm5
702	movdqa %xmm7,%xmm2
703	movl %edx,%ebp
704	movl %ecx,%ebx
705	.byte 102,15,56,0,253
706	movups (%edx),%xmm0
707	movups 16(%edx),%xmm1
708	leal 32(%edx),%edx
709	xorps %xmm0,%xmm2
710	L032enc1_loop_5:
711	.byte 102,15,56,220,209
712	decl %ecx
713	movups (%edx),%xmm1
714	leal 16(%edx),%edx
715	jnz L032enc1_loop_5
716	.byte 102,15,56,221,209
717	shll $4,%ebx
718	movl $16,%ecx
719	movups (%esi),%xmm6
720	paddq 16(%esp),%xmm7
721	leal 16(%esi),%esi
722	subl %ebx,%ecx
723	leal 32(%ebp,%ebx,1),%edx
724	movl %ecx,%ebx
725	jmp L033ccm64_dec_outer
726	.align 4,0x90
727	L033ccm64_dec_outer:
728	xorps %xmm2,%xmm6
729	movdqa %xmm7,%xmm2
730	movups %xmm6,(%edi)
731	leal 16(%edi),%edi
732	.byte 102,15,56,0,213
733	subl $1,%eax
734	jz L034ccm64_dec_break
735	movups (%ebp),%xmm0
736	movl %ebx,%ecx
737	movups 16(%ebp),%xmm1
738	xorps %xmm0,%xmm6
739	xorps %xmm0,%xmm2
740	xorps %xmm6,%xmm3
741	movups 32(%ebp),%xmm0
742	L035ccm64_dec2_loop:
743	.byte 102,15,56,220,209
744	.byte 102,15,56,220,217
745	movups (%edx,%ecx,1),%xmm1
746	addl $32,%ecx
747	.byte 102,15,56,220,208
748	.byte 102,15,56,220,216
749	movups -16(%edx,%ecx,1),%xmm0
750	jnz L035ccm64_dec2_loop
751	movups (%esi),%xmm6
752	paddq 16(%esp),%xmm7
753	.byte 102,15,56,220,209
754	.byte 102,15,56,220,217
755	.byte 102,15,56,221,208
756	.byte 102,15,56,221,216
757	leal 16(%esi),%esi
758	jmp L033ccm64_dec_outer
759	.align 4,0x90
760	L034ccm64_dec_break:
761	movl 240(%ebp),%ecx
762	movl %ebp,%edx
763	movups (%edx),%xmm0
764	movups 16(%edx),%xmm1
765	xorps %xmm0,%xmm6
766	leal 32(%edx),%edx
767	xorps %xmm6,%xmm3
768	L036enc1_loop_6:
769	.byte 102,15,56,220,217
770	decl %ecx
771	movups (%edx),%xmm1
772	leal 16(%edx),%edx
773	jnz L036enc1_loop_6
774	.byte 102,15,56,221,217
775	movl 48(%esp),%esp
776	movl 40(%esp),%edi
777	movups %xmm3,(%edi)
778	pxor %xmm0,%xmm0
779	pxor %xmm1,%xmm1
780	pxor %xmm2,%xmm2
781	pxor %xmm3,%xmm3
782	pxor %xmm4,%xmm4
783	pxor %xmm5,%xmm5
784	pxor %xmm6,%xmm6
785	pxor %xmm7,%xmm7
786	popl %edi
787	popl %esi
788	popl %ebx
789	popl %ebp
790	ret
791	.globl _aesni_ctr32_encrypt_blocks
792	.align 4
793	_aesni_ctr32_encrypt_blocks:
794	L_aesni_ctr32_encrypt_blocks_begin:
795	.byte 243,15,30,251
796	pushl %ebp
797	pushl %ebx
798	pushl %esi
799	pushl %edi
800	movl 20(%esp),%esi
801	movl 24(%esp),%edi
802	movl 28(%esp),%eax
803	movl 32(%esp),%edx
804	movl 36(%esp),%ebx
805	movl %esp,%ebp
806	subl $88,%esp
807	andl $-16,%esp
808	movl %ebp,80(%esp)
809	cmpl $1,%eax
810	je L037ctr32_one_shortcut
811	movdqu (%ebx),%xmm7
812	movl $202182159,(%esp)
813	movl $134810123,4(%esp)
814	movl $67438087,8(%esp)
815	movl $66051,12(%esp)
816	movl $6,%ecx
817	xorl %ebp,%ebp
818	movl %ecx,16(%esp)
819	movl %ecx,20(%esp)
820	movl %ecx,24(%esp)
821	movl %ebp,28(%esp)
822	.byte 102,15,58,22,251,3
823	.byte 102,15,58,34,253,3
824	movl 240(%edx),%ecx
825	bswap %ebx
826	pxor %xmm0,%xmm0
827	pxor %xmm1,%xmm1
828	movdqa (%esp),%xmm2
829	.byte 102,15,58,34,195,0
830	leal 3(%ebx),%ebp
831	.byte 102,15,58,34,205,0
832	incl %ebx
833	.byte 102,15,58,34,195,1
834	incl %ebp
835	.byte 102,15,58,34,205,1
836	incl %ebx
837	.byte 102,15,58,34,195,2
838	incl %ebp
839	.byte 102,15,58,34,205,2
840	movdqa %xmm0,48(%esp)
841	.byte 102,15,56,0,194
842	movdqu (%edx),%xmm6
843	movdqa %xmm1,64(%esp)
844	.byte 102,15,56,0,202
845	pshufd $192,%xmm0,%xmm2
846	pshufd $128,%xmm0,%xmm3
847	cmpl $6,%eax
848	jb L038ctr32_tail
849	pxor %xmm6,%xmm7
850	shll $4,%ecx
851	movl $16,%ebx
852	movdqa %xmm7,32(%esp)
853	movl %edx,%ebp
854	subl %ecx,%ebx
855	leal 32(%edx,%ecx,1),%edx
856	subl $6,%eax
857	jmp L039ctr32_loop6
858	.align 4,0x90
859	L039ctr32_loop6:
860	pshufd $64,%xmm0,%xmm4
861	movdqa 32(%esp),%xmm0
862	pshufd $192,%xmm1,%xmm5
863	pxor %xmm0,%xmm2
864	pshufd $128,%xmm1,%xmm6
865	pxor %xmm0,%xmm3
866	pshufd $64,%xmm1,%xmm7
867	movups 16(%ebp),%xmm1
868	pxor %xmm0,%xmm4
869	pxor %xmm0,%xmm5
870	.byte 102,15,56,220,209
871	pxor %xmm0,%xmm6
872	pxor %xmm0,%xmm7
873	.byte 102,15,56,220,217
874	movups 32(%ebp),%xmm0
875	movl %ebx,%ecx
876	.byte 102,15,56,220,225
877	.byte 102,15,56,220,233
878	.byte 102,15,56,220,241
879	.byte 102,15,56,220,249
880	call L_aesni_encrypt6_enter
881	movups (%esi),%xmm1
882	movups 16(%esi),%xmm0
883	xorps %xmm1,%xmm2
884	movups 32(%esi),%xmm1
885	xorps %xmm0,%xmm3
886	movups %xmm2,(%edi)
887	movdqa 16(%esp),%xmm0
888	xorps %xmm1,%xmm4
889	movdqa 64(%esp),%xmm1
890	movups %xmm3,16(%edi)
891	movups %xmm4,32(%edi)
892	paddd %xmm0,%xmm1
893	paddd 48(%esp),%xmm0
894	movdqa (%esp),%xmm2
895	movups 48(%esi),%xmm3
896	movups 64(%esi),%xmm4
897	xorps %xmm3,%xmm5
898	movups 80(%esi),%xmm3
899	leal 96(%esi),%esi
900	movdqa %xmm0,48(%esp)
901	.byte 102,15,56,0,194
902	xorps %xmm4,%xmm6
903	movups %xmm5,48(%edi)
904	xorps %xmm3,%xmm7
905	movdqa %xmm1,64(%esp)
906	.byte 102,15,56,0,202
907	movups %xmm6,64(%edi)
908	pshufd $192,%xmm0,%xmm2
909	movups %xmm7,80(%edi)
910	leal 96(%edi),%edi
911	pshufd $128,%xmm0,%xmm3
912	subl $6,%eax
913	jnc L039ctr32_loop6
914	addl $6,%eax
915	jz L040ctr32_ret
916	movdqu (%ebp),%xmm7
917	movl %ebp,%edx
918	pxor 32(%esp),%xmm7
919	movl 240(%ebp),%ecx
920	L038ctr32_tail:
921	por %xmm7,%xmm2
922	cmpl $2,%eax
923	jb L041ctr32_one
924	pshufd $64,%xmm0,%xmm4
925	por %xmm7,%xmm3
926	je L042ctr32_two
927	pshufd $192,%xmm1,%xmm5
928	por %xmm7,%xmm4
929	cmpl $4,%eax
930	jb L043ctr32_three
931	pshufd $128,%xmm1,%xmm6
932	por %xmm7,%xmm5
933	je L044ctr32_four
934	por %xmm7,%xmm6
935	call __aesni_encrypt6
936	movups (%esi),%xmm1
937	movups 16(%esi),%xmm0
938	xorps %xmm1,%xmm2
939	movups 32(%esi),%xmm1
940	xorps %xmm0,%xmm3
941	movups 48(%esi),%xmm0
942	xorps %xmm1,%xmm4
943	movups 64(%esi),%xmm1
944	xorps %xmm0,%xmm5
945	movups %xmm2,(%edi)
946	xorps %xmm1,%xmm6
947	movups %xmm3,16(%edi)
948	movups %xmm4,32(%edi)
949	movups %xmm5,48(%edi)
950	movups %xmm6,64(%edi)
951	jmp L040ctr32_ret
952	.align 4,0x90
953	L037ctr32_one_shortcut:
954	movups (%ebx),%xmm2
955	movl 240(%edx),%ecx
956	L041ctr32_one:
957	movups (%edx),%xmm0
958	movups 16(%edx),%xmm1
959	leal 32(%edx),%edx
960	xorps %xmm0,%xmm2
961	L045enc1_loop_7:
962	.byte 102,15,56,220,209
963	decl %ecx
964	movups (%edx),%xmm1
965	leal 16(%edx),%edx
966	jnz L045enc1_loop_7
967	.byte 102,15,56,221,209
968	movups (%esi),%xmm6
969	xorps %xmm2,%xmm6
970	movups %xmm6,(%edi)
971	jmp L040ctr32_ret
972	.align 4,0x90
973	L042ctr32_two:
974	call __aesni_encrypt2
975	movups (%esi),%xmm5
976	movups 16(%esi),%xmm6
977	xorps %xmm5,%xmm2
978	xorps %xmm6,%xmm3
979	movups %xmm2,(%edi)
980	movups %xmm3,16(%edi)
981	jmp L040ctr32_ret
982	.align 4,0x90
983	L043ctr32_three:
984	call __aesni_encrypt3
985	movups (%esi),%xmm5
986	movups 16(%esi),%xmm6
987	xorps %xmm5,%xmm2
988	movups 32(%esi),%xmm7
989	xorps %xmm6,%xmm3
990	movups %xmm2,(%edi)
991	xorps %xmm7,%xmm4
992	movups %xmm3,16(%edi)
993	movups %xmm4,32(%edi)
994	jmp L040ctr32_ret
995	.align 4,0x90
996	L044ctr32_four:
997	call __aesni_encrypt4
998	movups (%esi),%xmm6
999	movups 16(%esi),%xmm7
1000	movups 32(%esi),%xmm1
1001	xorps %xmm6,%xmm2
1002	movups 48(%esi),%xmm0
1003	xorps %xmm7,%xmm3
1004	movups %xmm2,(%edi)
1005	xorps %xmm1,%xmm4
1006	movups %xmm3,16(%edi)
1007	xorps %xmm0,%xmm5
1008	movups %xmm4,32(%edi)
1009	movups %xmm5,48(%edi)
1010	L040ctr32_ret:
1011	pxor %xmm0,%xmm0
1012	pxor %xmm1,%xmm1
1013	pxor %xmm2,%xmm2
1014	pxor %xmm3,%xmm3
1015	pxor %xmm4,%xmm4
1016	movdqa %xmm0,32(%esp)
1017	pxor %xmm5,%xmm5
1018	movdqa %xmm0,48(%esp)
1019	pxor %xmm6,%xmm6
1020	movdqa %xmm0,64(%esp)
1021	pxor %xmm7,%xmm7
1022	movl 80(%esp),%esp
1023	popl %edi
1024	popl %esi
1025	popl %ebx
1026	popl %ebp
1027	ret
1028	.globl _aesni_xts_encrypt
1029	.align 4
1030	_aesni_xts_encrypt:
1031	L_aesni_xts_encrypt_begin:
1032	.byte 243,15,30,251
1033	pushl %ebp
1034	pushl %ebx
1035	pushl %esi
1036	pushl %edi
1037	movl 36(%esp),%edx
1038	movl 40(%esp),%esi
1039	movl 240(%edx),%ecx
1040	movups (%esi),%xmm2
1041	movups (%edx),%xmm0
1042	movups 16(%edx),%xmm1
1043	leal 32(%edx),%edx
1044	xorps %xmm0,%xmm2
1045	L046enc1_loop_8:
1046	.byte 102,15,56,220,209
1047	decl %ecx
1048	movups (%edx),%xmm1
1049	leal 16(%edx),%edx
1050	jnz L046enc1_loop_8
1051	.byte 102,15,56,221,209
1052	movl 20(%esp),%esi
1053	movl 24(%esp),%edi
1054	movl 28(%esp),%eax
1055	movl 32(%esp),%edx
1056	movl %esp,%ebp
1057	subl $120,%esp
1058	movl 240(%edx),%ecx
1059	andl $-16,%esp
1060	movl $135,96(%esp)
1061	movl $0,100(%esp)
1062	movl $1,104(%esp)
1063	movl $0,108(%esp)
1064	movl %eax,112(%esp)
1065	movl %ebp,116(%esp)
1066	movdqa %xmm2,%xmm1
1067	pxor %xmm0,%xmm0
1068	movdqa 96(%esp),%xmm3
1069	pcmpgtd %xmm1,%xmm0
1070	andl $-16,%eax
1071	movl %edx,%ebp
1072	movl %ecx,%ebx
1073	subl $96,%eax
1074	jc L047xts_enc_short
1075	shll $4,%ecx
1076	movl $16,%ebx
1077	subl %ecx,%ebx
1078	leal 32(%edx,%ecx,1),%edx
1079	jmp L048xts_enc_loop6
1080	.align 4,0x90
1081	L048xts_enc_loop6:
1082	pshufd $19,%xmm0,%xmm2
1083	pxor %xmm0,%xmm0
1084	movdqa %xmm1,(%esp)
1085	paddq %xmm1,%xmm1
1086	pand %xmm3,%xmm2
1087	pcmpgtd %xmm1,%xmm0
1088	pxor %xmm2,%xmm1
1089	pshufd $19,%xmm0,%xmm2
1090	pxor %xmm0,%xmm0
1091	movdqa %xmm1,16(%esp)
1092	paddq %xmm1,%xmm1
1093	pand %xmm3,%xmm2
1094	pcmpgtd %xmm1,%xmm0
1095	pxor %xmm2,%xmm1
1096	pshufd $19,%xmm0,%xmm2
1097	pxor %xmm0,%xmm0
1098	movdqa %xmm1,32(%esp)
1099	paddq %xmm1,%xmm1
1100	pand %xmm3,%xmm2
1101	pcmpgtd %xmm1,%xmm0
1102	pxor %xmm2,%xmm1
1103	pshufd $19,%xmm0,%xmm2
1104	pxor %xmm0,%xmm0
1105	movdqa %xmm1,48(%esp)
1106	paddq %xmm1,%xmm1
1107	pand %xmm3,%xmm2
1108	pcmpgtd %xmm1,%xmm0
1109	pxor %xmm2,%xmm1
1110	pshufd $19,%xmm0,%xmm7
1111	movdqa %xmm1,64(%esp)
1112	paddq %xmm1,%xmm1
1113	movups (%ebp),%xmm0
1114	pand %xmm3,%xmm7
1115	movups (%esi),%xmm2
1116	pxor %xmm1,%xmm7
1117	movl %ebx,%ecx
1118	movdqu 16(%esi),%xmm3
1119	xorps %xmm0,%xmm2
1120	movdqu 32(%esi),%xmm4
1121	pxor %xmm0,%xmm3
1122	movdqu 48(%esi),%xmm5
1123	pxor %xmm0,%xmm4
1124	movdqu 64(%esi),%xmm6
1125	pxor %xmm0,%xmm5
1126	movdqu 80(%esi),%xmm1
1127	pxor %xmm0,%xmm6
1128	leal 96(%esi),%esi
1129	pxor (%esp),%xmm2
1130	movdqa %xmm7,80(%esp)
1131	pxor %xmm1,%xmm7
1132	movups 16(%ebp),%xmm1
1133	pxor 16(%esp),%xmm3
1134	pxor 32(%esp),%xmm4
1135	.byte 102,15,56,220,209
1136	pxor 48(%esp),%xmm5
1137	pxor 64(%esp),%xmm6
1138	.byte 102,15,56,220,217
1139	pxor %xmm0,%xmm7
1140	movups 32(%ebp),%xmm0
1141	.byte 102,15,56,220,225
1142	.byte 102,15,56,220,233
1143	.byte 102,15,56,220,241
1144	.byte 102,15,56,220,249
1145	call L_aesni_encrypt6_enter
1146	movdqa 80(%esp),%xmm1
1147	pxor %xmm0,%xmm0
1148	xorps (%esp),%xmm2
1149	pcmpgtd %xmm1,%xmm0
1150	xorps 16(%esp),%xmm3
1151	movups %xmm2,(%edi)
1152	xorps 32(%esp),%xmm4
1153	movups %xmm3,16(%edi)
1154	xorps 48(%esp),%xmm5
1155	movups %xmm4,32(%edi)
1156	xorps 64(%esp),%xmm6
1157	movups %xmm5,48(%edi)
1158	xorps %xmm1,%xmm7
1159	movups %xmm6,64(%edi)
1160	pshufd $19,%xmm0,%xmm2
1161	movups %xmm7,80(%edi)
1162	leal 96(%edi),%edi
1163	movdqa 96(%esp),%xmm3
1164	pxor %xmm0,%xmm0
1165	paddq %xmm1,%xmm1
1166	pand %xmm3,%xmm2
1167	pcmpgtd %xmm1,%xmm0
1168	pxor %xmm2,%xmm1
1169	subl $96,%eax
1170	jnc L048xts_enc_loop6
1171	movl 240(%ebp),%ecx
1172	movl %ebp,%edx
1173	movl %ecx,%ebx
1174	L047xts_enc_short:
1175	addl $96,%eax
1176	jz L049xts_enc_done6x
1177	movdqa %xmm1,%xmm5
1178	cmpl $32,%eax
1179	jb L050xts_enc_one
1180	pshufd $19,%xmm0,%xmm2
1181	pxor %xmm0,%xmm0
1182	paddq %xmm1,%xmm1
1183	pand %xmm3,%xmm2
1184	pcmpgtd %xmm1,%xmm0
1185	pxor %xmm2,%xmm1
1186	je L051xts_enc_two
1187	pshufd $19,%xmm0,%xmm2
1188	pxor %xmm0,%xmm0
1189	movdqa %xmm1,%xmm6
1190	paddq %xmm1,%xmm1
1191	pand %xmm3,%xmm2
1192	pcmpgtd %xmm1,%xmm0
1193	pxor %xmm2,%xmm1
1194	cmpl $64,%eax
1195	jb L052xts_enc_three
1196	pshufd $19,%xmm0,%xmm2
1197	pxor %xmm0,%xmm0
1198	movdqa %xmm1,%xmm7
1199	paddq %xmm1,%xmm1
1200	pand %xmm3,%xmm2
1201	pcmpgtd %xmm1,%xmm0
1202	pxor %xmm2,%xmm1
1203	movdqa %xmm5,(%esp)
1204	movdqa %xmm6,16(%esp)
1205	je L053xts_enc_four
1206	movdqa %xmm7,32(%esp)
1207	pshufd $19,%xmm0,%xmm7
1208	movdqa %xmm1,48(%esp)
1209	paddq %xmm1,%xmm1
1210	pand %xmm3,%xmm7
1211	pxor %xmm1,%xmm7
1212	movdqu (%esi),%xmm2
1213	movdqu 16(%esi),%xmm3
1214	movdqu 32(%esi),%xmm4
1215	pxor (%esp),%xmm2
1216	movdqu 48(%esi),%xmm5
1217	pxor 16(%esp),%xmm3
1218	movdqu 64(%esi),%xmm6
1219	pxor 32(%esp),%xmm4
1220	leal 80(%esi),%esi
1221	pxor 48(%esp),%xmm5
1222	movdqa %xmm7,64(%esp)
1223	pxor %xmm7,%xmm6
1224	call __aesni_encrypt6
1225	movaps 64(%esp),%xmm1
1226	xorps (%esp),%xmm2
1227	xorps 16(%esp),%xmm3
1228	xorps 32(%esp),%xmm4
1229	movups %xmm2,(%edi)
1230	xorps 48(%esp),%xmm5
1231	movups %xmm3,16(%edi)
1232	xorps %xmm1,%xmm6
1233	movups %xmm4,32(%edi)
1234	movups %xmm5,48(%edi)
1235	movups %xmm6,64(%edi)
1236	leal 80(%edi),%edi
1237	jmp L054xts_enc_done
1238	.align 4,0x90
1239	L050xts_enc_one:
1240	movups (%esi),%xmm2
1241	leal 16(%esi),%esi
1242	xorps %xmm5,%xmm2
1243	movups (%edx),%xmm0
1244	movups 16(%edx),%xmm1
1245	leal 32(%edx),%edx
1246	xorps %xmm0,%xmm2
1247	L055enc1_loop_9:
1248	.byte 102,15,56,220,209
1249	decl %ecx
1250	movups (%edx),%xmm1
1251	leal 16(%edx),%edx
1252	jnz L055enc1_loop_9
1253	.byte 102,15,56,221,209
1254	xorps %xmm5,%xmm2
1255	movups %xmm2,(%edi)
1256	leal 16(%edi),%edi
1257	movdqa %xmm5,%xmm1
1258	jmp L054xts_enc_done
1259	.align 4,0x90
1260	L051xts_enc_two:
1261	movaps %xmm1,%xmm6
1262	movups (%esi),%xmm2
1263	movups 16(%esi),%xmm3
1264	leal 32(%esi),%esi
1265	xorps %xmm5,%xmm2
1266	xorps %xmm6,%xmm3
1267	call __aesni_encrypt2
1268	xorps %xmm5,%xmm2
1269	xorps %xmm6,%xmm3
1270	movups %xmm2,(%edi)
1271	movups %xmm3,16(%edi)
1272	leal 32(%edi),%edi
1273	movdqa %xmm6,%xmm1
1274	jmp L054xts_enc_done
1275	.align 4,0x90
1276	L052xts_enc_three:
1277	movaps %xmm1,%xmm7
1278	movups (%esi),%xmm2
1279	movups 16(%esi),%xmm3
1280	movups 32(%esi),%xmm4
1281	leal 48(%esi),%esi
1282	xorps %xmm5,%xmm2
1283	xorps %xmm6,%xmm3
1284	xorps %xmm7,%xmm4
1285	call __aesni_encrypt3
1286	xorps %xmm5,%xmm2
1287	xorps %xmm6,%xmm3
1288	xorps %xmm7,%xmm4
1289	movups %xmm2,(%edi)
1290	movups %xmm3,16(%edi)
1291	movups %xmm4,32(%edi)
1292	leal 48(%edi),%edi
1293	movdqa %xmm7,%xmm1
1294	jmp L054xts_enc_done
1295	.align 4,0x90
1296	L053xts_enc_four:
1297	movaps %xmm1,%xmm6
1298	movups (%esi),%xmm2
1299	movups 16(%esi),%xmm3
1300	movups 32(%esi),%xmm4
1301	xorps (%esp),%xmm2
1302	movups 48(%esi),%xmm5
1303	leal 64(%esi),%esi
1304	xorps 16(%esp),%xmm3
1305	xorps %xmm7,%xmm4
1306	xorps %xmm6,%xmm5
1307	call __aesni_encrypt4
1308	xorps (%esp),%xmm2
1309	xorps 16(%esp),%xmm3
1310	xorps %xmm7,%xmm4
1311	movups %xmm2,(%edi)
1312	xorps %xmm6,%xmm5
1313	movups %xmm3,16(%edi)
1314	movups %xmm4,32(%edi)
1315	movups %xmm5,48(%edi)
1316	leal 64(%edi),%edi
1317	movdqa %xmm6,%xmm1
1318	jmp L054xts_enc_done
1319	.align 4,0x90
1320	L049xts_enc_done6x:
1321	movl 112(%esp),%eax
1322	andl $15,%eax
1323	jz L056xts_enc_ret
1324	movdqa %xmm1,%xmm5
1325	movl %eax,112(%esp)
1326	jmp L057xts_enc_steal
1327	.align 4,0x90
1328	L054xts_enc_done:
1329	movl 112(%esp),%eax
1330	pxor %xmm0,%xmm0
1331	andl $15,%eax
1332	jz L056xts_enc_ret
1333	pcmpgtd %xmm1,%xmm0
1334	movl %eax,112(%esp)
1335	pshufd $19,%xmm0,%xmm5
1336	paddq %xmm1,%xmm1
1337	pand 96(%esp),%xmm5
1338	pxor %xmm1,%xmm5
1339	L057xts_enc_steal:
1340	movzbl (%esi),%ecx
1341	movzbl -16(%edi),%edx
1342	leal 1(%esi),%esi
1343	movb %cl,-16(%edi)
1344	movb %dl,(%edi)
1345	leal 1(%edi),%edi
1346	subl $1,%eax
1347	jnz L057xts_enc_steal
1348	subl 112(%esp),%edi
1349	movl %ebp,%edx
1350	movl %ebx,%ecx
1351	movups -16(%edi),%xmm2
1352	xorps %xmm5,%xmm2
1353	movups (%edx),%xmm0
1354	movups 16(%edx),%xmm1
1355	leal 32(%edx),%edx
1356	xorps %xmm0,%xmm2
1357	L058enc1_loop_10:
1358	.byte 102,15,56,220,209
1359	decl %ecx
1360	movups (%edx),%xmm1
1361	leal 16(%edx),%edx
1362	jnz L058enc1_loop_10
1363	.byte 102,15,56,221,209
1364	xorps %xmm5,%xmm2
1365	movups %xmm2,-16(%edi)
1366	L056xts_enc_ret:
1367	pxor %xmm0,%xmm0
1368	pxor %xmm1,%xmm1
1369	pxor %xmm2,%xmm2
1370	movdqa %xmm0,(%esp)
1371	pxor %xmm3,%xmm3
1372	movdqa %xmm0,16(%esp)
1373	pxor %xmm4,%xmm4
1374	movdqa %xmm0,32(%esp)
1375	pxor %xmm5,%xmm5
1376	movdqa %xmm0,48(%esp)
1377	pxor %xmm6,%xmm6
1378	movdqa %xmm0,64(%esp)
1379	pxor %xmm7,%xmm7
1380	movdqa %xmm0,80(%esp)
1381	movl 116(%esp),%esp
1382	popl %edi
1383	popl %esi
1384	popl %ebx
1385	popl %ebp
1386	ret
1387	.globl _aesni_xts_decrypt
1388	.align 4
1389	_aesni_xts_decrypt:
1390	L_aesni_xts_decrypt_begin:
1391	.byte 243,15,30,251
1392	pushl %ebp
1393	pushl %ebx
1394	pushl %esi
1395	pushl %edi
1396	movl 36(%esp),%edx
1397	movl 40(%esp),%esi
1398	movl 240(%edx),%ecx
1399	movups (%esi),%xmm2
1400	movups (%edx),%xmm0
1401	movups 16(%edx),%xmm1
1402	leal 32(%edx),%edx
1403	xorps %xmm0,%xmm2
1404	L059enc1_loop_11:
1405	.byte 102,15,56,220,209
1406	decl %ecx
1407	movups (%edx),%xmm1
1408	leal 16(%edx),%edx
1409	jnz L059enc1_loop_11
1410	.byte 102,15,56,221,209
1411	movl 20(%esp),%esi
1412	movl 24(%esp),%edi
1413	movl 28(%esp),%eax
1414	movl 32(%esp),%edx
1415	movl %esp,%ebp
1416	subl $120,%esp
1417	andl $-16,%esp
1418	xorl %ebx,%ebx
1419	testl $15,%eax
1420	setnz %bl
1421	shll $4,%ebx
1422	subl %ebx,%eax
1423	movl $135,96(%esp)
1424	movl $0,100(%esp)
1425	movl $1,104(%esp)
1426	movl $0,108(%esp)
1427	movl %eax,112(%esp)
1428	movl %ebp,116(%esp)
1429	movl 240(%edx),%ecx
1430	movl %edx,%ebp
1431	movl %ecx,%ebx
1432	movdqa %xmm2,%xmm1
1433	pxor %xmm0,%xmm0
1434	movdqa 96(%esp),%xmm3
1435	pcmpgtd %xmm1,%xmm0
1436	andl $-16,%eax
1437	subl $96,%eax
1438	jc L060xts_dec_short
1439	shll $4,%ecx
1440	movl $16,%ebx
1441	subl %ecx,%ebx
1442	leal 32(%edx,%ecx,1),%edx
1443	jmp L061xts_dec_loop6
1444	.align 4,0x90
1445	L061xts_dec_loop6:
1446	pshufd $19,%xmm0,%xmm2
1447	pxor %xmm0,%xmm0
1448	movdqa %xmm1,(%esp)
1449	paddq %xmm1,%xmm1
1450	pand %xmm3,%xmm2
1451	pcmpgtd %xmm1,%xmm0
1452	pxor %xmm2,%xmm1
1453	pshufd $19,%xmm0,%xmm2
1454	pxor %xmm0,%xmm0
1455	movdqa %xmm1,16(%esp)
1456	paddq %xmm1,%xmm1
1457	pand %xmm3,%xmm2
1458	pcmpgtd %xmm1,%xmm0
1459	pxor %xmm2,%xmm1
1460	pshufd $19,%xmm0,%xmm2
1461	pxor %xmm0,%xmm0
1462	movdqa %xmm1,32(%esp)
1463	paddq %xmm1,%xmm1
1464	pand %xmm3,%xmm2
1465	pcmpgtd %xmm1,%xmm0
1466	pxor %xmm2,%xmm1
1467	pshufd $19,%xmm0,%xmm2
1468	pxor %xmm0,%xmm0
1469	movdqa %xmm1,48(%esp)
1470	paddq %xmm1,%xmm1
1471	pand %xmm3,%xmm2
1472	pcmpgtd %xmm1,%xmm0
1473	pxor %xmm2,%xmm1
1474	pshufd $19,%xmm0,%xmm7
1475	movdqa %xmm1,64(%esp)
1476	paddq %xmm1,%xmm1
1477	movups (%ebp),%xmm0
1478	pand %xmm3,%xmm7
1479	movups (%esi),%xmm2
1480	pxor %xmm1,%xmm7
1481	movl %ebx,%ecx
1482	movdqu 16(%esi),%xmm3
1483	xorps %xmm0,%xmm2
1484	movdqu 32(%esi),%xmm4
1485	pxor %xmm0,%xmm3
1486	movdqu 48(%esi),%xmm5
1487	pxor %xmm0,%xmm4
1488	movdqu 64(%esi),%xmm6
1489	pxor %xmm0,%xmm5
1490	movdqu 80(%esi),%xmm1
1491	pxor %xmm0,%xmm6
1492	leal 96(%esi),%esi
1493	pxor (%esp),%xmm2
1494	movdqa %xmm7,80(%esp)
1495	pxor %xmm1,%xmm7
1496	movups 16(%ebp),%xmm1
1497	pxor 16(%esp),%xmm3
1498	pxor 32(%esp),%xmm4
1499	.byte 102,15,56,222,209
1500	pxor 48(%esp),%xmm5
1501	pxor 64(%esp),%xmm6
1502	.byte 102,15,56,222,217
1503	pxor %xmm0,%xmm7
1504	movups 32(%ebp),%xmm0
1505	.byte 102,15,56,222,225
1506	.byte 102,15,56,222,233
1507	.byte 102,15,56,222,241
1508	.byte 102,15,56,222,249
1509	call L_aesni_decrypt6_enter
1510	movdqa 80(%esp),%xmm1
1511	pxor %xmm0,%xmm0
1512	xorps (%esp),%xmm2
1513	pcmpgtd %xmm1,%xmm0
1514	xorps 16(%esp),%xmm3
1515	movups %xmm2,(%edi)
1516	xorps 32(%esp),%xmm4
1517	movups %xmm3,16(%edi)
1518	xorps 48(%esp),%xmm5
1519	movups %xmm4,32(%edi)
1520	xorps 64(%esp),%xmm6
1521	movups %xmm5,48(%edi)
1522	xorps %xmm1,%xmm7
1523	movups %xmm6,64(%edi)
1524	pshufd $19,%xmm0,%xmm2
1525	movups %xmm7,80(%edi)
1526	leal 96(%edi),%edi
1527	movdqa 96(%esp),%xmm3
1528	pxor %xmm0,%xmm0
1529	paddq %xmm1,%xmm1
1530	pand %xmm3,%xmm2
1531	pcmpgtd %xmm1,%xmm0
1532	pxor %xmm2,%xmm1
1533	subl $96,%eax
1534	jnc L061xts_dec_loop6
1535	movl 240(%ebp),%ecx
1536	movl %ebp,%edx
1537	movl %ecx,%ebx
1538	L060xts_dec_short:
1539	addl $96,%eax
1540	jz L062xts_dec_done6x
1541	movdqa %xmm1,%xmm5
1542	cmpl $32,%eax
1543	jb L063xts_dec_one
1544	pshufd $19,%xmm0,%xmm2
1545	pxor %xmm0,%xmm0
1546	paddq %xmm1,%xmm1
1547	pand %xmm3,%xmm2
1548	pcmpgtd %xmm1,%xmm0
1549	pxor %xmm2,%xmm1
1550	je L064xts_dec_two
1551	pshufd $19,%xmm0,%xmm2
1552	pxor %xmm0,%xmm0
1553	movdqa %xmm1,%xmm6
1554	paddq %xmm1,%xmm1
1555	pand %xmm3,%xmm2
1556	pcmpgtd %xmm1,%xmm0
1557	pxor %xmm2,%xmm1
1558	cmpl $64,%eax
1559	jb L065xts_dec_three
1560	pshufd $19,%xmm0,%xmm2
1561	pxor %xmm0,%xmm0
1562	movdqa %xmm1,%xmm7
1563	paddq %xmm1,%xmm1
1564	pand %xmm3,%xmm2
1565	pcmpgtd %xmm1,%xmm0
1566	pxor %xmm2,%xmm1
1567	movdqa %xmm5,(%esp)
1568	movdqa %xmm6,16(%esp)
1569	je L066xts_dec_four
1570	movdqa %xmm7,32(%esp)
1571	pshufd $19,%xmm0,%xmm7
1572	movdqa %xmm1,48(%esp)
1573	paddq %xmm1,%xmm1
1574	pand %xmm3,%xmm7
1575	pxor %xmm1,%xmm7
1576	movdqu (%esi),%xmm2
1577	movdqu 16(%esi),%xmm3
1578	movdqu 32(%esi),%xmm4
1579	pxor (%esp),%xmm2
1580	movdqu 48(%esi),%xmm5
1581	pxor 16(%esp),%xmm3
1582	movdqu 64(%esi),%xmm6
1583	pxor 32(%esp),%xmm4
1584	leal 80(%esi),%esi
1585	pxor 48(%esp),%xmm5
1586	movdqa %xmm7,64(%esp)
1587	pxor %xmm7,%xmm6
1588	call __aesni_decrypt6
1589	movaps 64(%esp),%xmm1
1590	xorps (%esp),%xmm2
1591	xorps 16(%esp),%xmm3
1592	xorps 32(%esp),%xmm4
1593	movups %xmm2,(%edi)
1594	xorps 48(%esp),%xmm5
1595	movups %xmm3,16(%edi)
1596	xorps %xmm1,%xmm6
1597	movups %xmm4,32(%edi)
1598	movups %xmm5,48(%edi)
1599	movups %xmm6,64(%edi)
1600	leal 80(%edi),%edi
1601	jmp L067xts_dec_done
1602	.align 4,0x90
1603	L063xts_dec_one:
1604	movups (%esi),%xmm2
1605	leal 16(%esi),%esi
1606	xorps %xmm5,%xmm2
1607	movups (%edx),%xmm0
1608	movups 16(%edx),%xmm1
1609	leal 32(%edx),%edx
1610	xorps %xmm0,%xmm2
1611	L068dec1_loop_12:
1612	.byte 102,15,56,222,209
1613	decl %ecx
1614	movups (%edx),%xmm1
1615	leal 16(%edx),%edx
1616	jnz L068dec1_loop_12
1617	.byte 102,15,56,223,209
1618	xorps %xmm5,%xmm2
1619	movups %xmm2,(%edi)
1620	leal 16(%edi),%edi
1621	movdqa %xmm5,%xmm1
1622	jmp L067xts_dec_done
1623	.align 4,0x90
1624	L064xts_dec_two:
1625	movaps %xmm1,%xmm6
1626	movups (%esi),%xmm2
1627	movups 16(%esi),%xmm3
1628	leal 32(%esi),%esi
1629	xorps %xmm5,%xmm2
1630	xorps %xmm6,%xmm3
1631	call __aesni_decrypt2
1632	xorps %xmm5,%xmm2
1633	xorps %xmm6,%xmm3
1634	movups %xmm2,(%edi)
1635	movups %xmm3,16(%edi)
1636	leal 32(%edi),%edi
1637	movdqa %xmm6,%xmm1
1638	jmp L067xts_dec_done
1639	.align 4,0x90
1640	L065xts_dec_three:
1641	movaps %xmm1,%xmm7
1642	movups (%esi),%xmm2
1643	movups 16(%esi),%xmm3
1644	movups 32(%esi),%xmm4
1645	leal 48(%esi),%esi
1646	xorps %xmm5,%xmm2
1647	xorps %xmm6,%xmm3
1648	xorps %xmm7,%xmm4
1649	call __aesni_decrypt3
1650	xorps %xmm5,%xmm2
1651	xorps %xmm6,%xmm3
1652	xorps %xmm7,%xmm4
1653	movups %xmm2,(%edi)
1654	movups %xmm3,16(%edi)
1655	movups %xmm4,32(%edi)
1656	leal 48(%edi),%edi
1657	movdqa %xmm7,%xmm1
1658	jmp L067xts_dec_done
1659	.align 4,0x90
1660	L066xts_dec_four:
1661	movaps %xmm1,%xmm6
1662	movups (%esi),%xmm2
1663	movups 16(%esi),%xmm3
1664	movups 32(%esi),%xmm4
1665	xorps (%esp),%xmm2
1666	movups 48(%esi),%xmm5
1667	leal 64(%esi),%esi
1668	xorps 16(%esp),%xmm3
1669	xorps %xmm7,%xmm4
1670	xorps %xmm6,%xmm5
1671	call __aesni_decrypt4
1672	xorps (%esp),%xmm2
1673	xorps 16(%esp),%xmm3
1674	xorps %xmm7,%xmm4
1675	movups %xmm2,(%edi)
1676	xorps %xmm6,%xmm5
1677	movups %xmm3,16(%edi)
1678	movups %xmm4,32(%edi)
1679	movups %xmm5,48(%edi)
1680	leal 64(%edi),%edi
1681	movdqa %xmm6,%xmm1
1682	jmp L067xts_dec_done
1683	.align 4,0x90
1684	L062xts_dec_done6x:
1685	movl 112(%esp),%eax
1686	andl $15,%eax
1687	jz L069xts_dec_ret
1688	movl %eax,112(%esp)
1689	jmp L070xts_dec_only_one_more
1690	.align 4,0x90
1691	L067xts_dec_done:
1692	movl 112(%esp),%eax
1693	pxor %xmm0,%xmm0
1694	andl $15,%eax
1695	jz L069xts_dec_ret
1696	pcmpgtd %xmm1,%xmm0
1697	movl %eax,112(%esp)
1698	pshufd $19,%xmm0,%xmm2
1699	pxor %xmm0,%xmm0
1700	movdqa 96(%esp),%xmm3
1701	paddq %xmm1,%xmm1
1702	pand %xmm3,%xmm2
1703	pcmpgtd %xmm1,%xmm0
1704	pxor %xmm2,%xmm1
1705	L070xts_dec_only_one_more:
1706	pshufd $19,%xmm0,%xmm5
1707	movdqa %xmm1,%xmm6
1708	paddq %xmm1,%xmm1
1709	pand %xmm3,%xmm5
1710	pxor %xmm1,%xmm5
1711	movl %ebp,%edx
1712	movl %ebx,%ecx
1713	movups (%esi),%xmm2
1714	xorps %xmm5,%xmm2
1715	movups (%edx),%xmm0
1716	movups 16(%edx),%xmm1
1717	leal 32(%edx),%edx
1718	xorps %xmm0,%xmm2
1719	L071dec1_loop_13:
1720	.byte 102,15,56,222,209
1721	decl %ecx
1722	movups (%edx),%xmm1
1723	leal 16(%edx),%edx
1724	jnz L071dec1_loop_13
1725	.byte 102,15,56,223,209
1726	xorps %xmm5,%xmm2
1727	movups %xmm2,(%edi)
1728	L072xts_dec_steal:
1729	movzbl 16(%esi),%ecx
1730	movzbl (%edi),%edx
1731	leal 1(%esi),%esi
1732	movb %cl,(%edi)
1733	movb %dl,16(%edi)
1734	leal 1(%edi),%edi
1735	subl $1,%eax
1736	jnz L072xts_dec_steal
1737	subl 112(%esp),%edi
1738	movl %ebp,%edx
1739	movl %ebx,%ecx
1740	movups (%edi),%xmm2
1741	xorps %xmm6,%xmm2
1742	movups (%edx),%xmm0
1743	movups 16(%edx),%xmm1
1744	leal 32(%edx),%edx
1745	xorps %xmm0,%xmm2
1746	L073dec1_loop_14:
1747	.byte 102,15,56,222,209
1748	decl %ecx
1749	movups (%edx),%xmm1
1750	leal 16(%edx),%edx
1751	jnz L073dec1_loop_14
1752	.byte 102,15,56,223,209
1753	xorps %xmm6,%xmm2
1754	movups %xmm2,(%edi)
1755	L069xts_dec_ret:
1756	pxor %xmm0,%xmm0
1757	pxor %xmm1,%xmm1
1758	pxor %xmm2,%xmm2
1759	movdqa %xmm0,(%esp)
1760	pxor %xmm3,%xmm3
1761	movdqa %xmm0,16(%esp)
1762	pxor %xmm4,%xmm4
1763	movdqa %xmm0,32(%esp)
1764	pxor %xmm5,%xmm5
1765	movdqa %xmm0,48(%esp)
1766	pxor %xmm6,%xmm6
1767	movdqa %xmm0,64(%esp)
1768	pxor %xmm7,%xmm7
1769	movdqa %xmm0,80(%esp)
1770	movl 116(%esp),%esp
1771	popl %edi
1772	popl %esi
1773	popl %ebx
1774	popl %ebp
1775	ret
1776	.globl _aesni_ocb_encrypt
1777	.align 4
1778	_aesni_ocb_encrypt:
1779	L_aesni_ocb_encrypt_begin:
1780	.byte 243,15,30,251
1781	pushl %ebp
1782	pushl %ebx
1783	pushl %esi
1784	pushl %edi
1785	movl 40(%esp),%ecx
1786	movl 48(%esp),%ebx
1787	movl 20(%esp),%esi
1788	movl 24(%esp),%edi
1789	movl 28(%esp),%eax
1790	movl 32(%esp),%edx
1791	movdqu (%ecx),%xmm0
1792	movl 36(%esp),%ebp
1793	movdqu (%ebx),%xmm1
1794	movl 44(%esp),%ebx
1795	movl %esp,%ecx
1796	subl $132,%esp
1797	andl $-16,%esp
1798	subl %esi,%edi
1799	shll $4,%eax
1800	leal -96(%esi,%eax,1),%eax
1801	movl %edi,120(%esp)
1802	movl %eax,124(%esp)
1803	movl %ecx,128(%esp)
1804	movl 240(%edx),%ecx
1805	testl $1,%ebp
1806	jnz L074odd
1807	bsfl %ebp,%eax
1808	addl $1,%ebp
1809	shll $4,%eax
1810	movdqu (%ebx,%eax,1),%xmm7
1811	movl %edx,%eax
1812	movdqu (%esi),%xmm2
1813	leal 16(%esi),%esi
1814	pxor %xmm0,%xmm7
1815	pxor %xmm2,%xmm1
1816	pxor %xmm7,%xmm2
1817	movdqa %xmm1,%xmm6
1818	movups (%edx),%xmm0
1819	movups 16(%edx),%xmm1
1820	leal 32(%edx),%edx
1821	xorps %xmm0,%xmm2
1822	L075enc1_loop_15:
1823	.byte 102,15,56,220,209
1824	decl %ecx
1825	movups (%edx),%xmm1
1826	leal 16(%edx),%edx
1827	jnz L075enc1_loop_15
1828	.byte 102,15,56,221,209
1829	xorps %xmm7,%xmm2
1830	movdqa %xmm7,%xmm0
1831	movdqa %xmm6,%xmm1
1832	movups %xmm2,-16(%edi,%esi,1)
1833	movl 240(%eax),%ecx
1834	movl %eax,%edx
1835	movl 124(%esp),%eax
1836	L074odd:
1837	shll $4,%ecx
1838	movl $16,%edi
1839	subl %ecx,%edi
1840	movl %edx,112(%esp)
1841	leal 32(%edx,%ecx,1),%edx
1842	movl %edi,116(%esp)
1843	cmpl %eax,%esi
1844	ja L076short
1845	jmp L077grandloop
1846	.align 5,0x90
1847	L077grandloop:
1848	leal 1(%ebp),%ecx
1849	leal 3(%ebp),%eax
1850	leal 5(%ebp),%edi
1851	addl $6,%ebp
1852	bsfl %ecx,%ecx
1853	bsfl %eax,%eax
1854	bsfl %edi,%edi
1855	shll $4,%ecx
1856	shll $4,%eax
1857	shll $4,%edi
1858	movdqu (%ebx),%xmm2
1859	movdqu (%ebx,%ecx,1),%xmm3
1860	movl 116(%esp),%ecx
1861	movdqa %xmm2,%xmm4
1862	movdqu (%ebx,%eax,1),%xmm5
1863	movdqa %xmm2,%xmm6
1864	movdqu (%ebx,%edi,1),%xmm7
1865	pxor %xmm0,%xmm2
1866	pxor %xmm2,%xmm3
1867	movdqa %xmm2,(%esp)
1868	pxor %xmm3,%xmm4
1869	movdqa %xmm3,16(%esp)
1870	pxor %xmm4,%xmm5
1871	movdqa %xmm4,32(%esp)
1872	pxor %xmm5,%xmm6
1873	movdqa %xmm5,48(%esp)
1874	pxor %xmm6,%xmm7
1875	movdqa %xmm6,64(%esp)
1876	movdqa %xmm7,80(%esp)
1877	movups -48(%edx,%ecx,1),%xmm0
1878	movdqu (%esi),%xmm2
1879	movdqu 16(%esi),%xmm3
1880	movdqu 32(%esi),%xmm4
1881	movdqu 48(%esi),%xmm5
1882	movdqu 64(%esi),%xmm6
1883	movdqu 80(%esi),%xmm7
1884	leal 96(%esi),%esi
1885	pxor %xmm2,%xmm1
1886	pxor %xmm0,%xmm2
1887	pxor %xmm3,%xmm1
1888	pxor %xmm0,%xmm3
1889	pxor %xmm4,%xmm1
1890	pxor %xmm0,%xmm4
1891	pxor %xmm5,%xmm1
1892	pxor %xmm0,%xmm5
1893	pxor %xmm6,%xmm1
1894	pxor %xmm0,%xmm6
1895	pxor %xmm7,%xmm1
1896	pxor %xmm0,%xmm7
1897	movdqa %xmm1,96(%esp)
1898	movups -32(%edx,%ecx,1),%xmm1
1899	pxor (%esp),%xmm2
1900	pxor 16(%esp),%xmm3
1901	pxor 32(%esp),%xmm4
1902	pxor 48(%esp),%xmm5
1903	pxor 64(%esp),%xmm6
1904	pxor 80(%esp),%xmm7
1905	movups -16(%edx,%ecx,1),%xmm0
1906	.byte 102,15,56,220,209
1907	.byte 102,15,56,220,217
1908	.byte 102,15,56,220,225
1909	.byte 102,15,56,220,233
1910	.byte 102,15,56,220,241
1911	.byte 102,15,56,220,249
1912	movl 120(%esp),%edi
1913	movl 124(%esp),%eax
1914	call L_aesni_encrypt6_enter
1915	movdqa 80(%esp),%xmm0
1916	pxor (%esp),%xmm2
1917	pxor 16(%esp),%xmm3
1918	pxor 32(%esp),%xmm4
1919	pxor 48(%esp),%xmm5
1920	pxor 64(%esp),%xmm6
1921	pxor %xmm0,%xmm7
1922	movdqa 96(%esp),%xmm1
1923	movdqu %xmm2,-96(%edi,%esi,1)
1924	movdqu %xmm3,-80(%edi,%esi,1)
1925	movdqu %xmm4,-64(%edi,%esi,1)
1926	movdqu %xmm5,-48(%edi,%esi,1)
1927	movdqu %xmm6,-32(%edi,%esi,1)
1928	movdqu %xmm7,-16(%edi,%esi,1)
1929	cmpl %eax,%esi
1930	jb L077grandloop
1931	L076short:
1932	addl $96,%eax
1933	subl %esi,%eax
1934	jz L078done
1935	cmpl $32,%eax
1936	jb L079one
1937	je L080two
1938	cmpl $64,%eax
1939	jb L081three
1940	je L082four
1941	leal 1(%ebp),%ecx
1942	leal 3(%ebp),%eax
1943	bsfl %ecx,%ecx
1944	bsfl %eax,%eax
1945	shll $4,%ecx
1946	shll $4,%eax
1947	movdqu (%ebx),%xmm2
1948	movdqu (%ebx,%ecx,1),%xmm3
1949	movl 116(%esp),%ecx
1950	movdqa %xmm2,%xmm4
1951	movdqu (%ebx,%eax,1),%xmm5
1952	movdqa %xmm2,%xmm6
1953	pxor %xmm0,%xmm2
1954	pxor %xmm2,%xmm3
1955	movdqa %xmm2,(%esp)
1956	pxor %xmm3,%xmm4
1957	movdqa %xmm3,16(%esp)
1958	pxor %xmm4,%xmm5
1959	movdqa %xmm4,32(%esp)
1960	pxor %xmm5,%xmm6
1961	movdqa %xmm5,48(%esp)
1962	pxor %xmm6,%xmm7
1963	movdqa %xmm6,64(%esp)
1964	movups -48(%edx,%ecx,1),%xmm0
1965	movdqu (%esi),%xmm2
1966	movdqu 16(%esi),%xmm3
1967	movdqu 32(%esi),%xmm4
1968	movdqu 48(%esi),%xmm5
1969	movdqu 64(%esi),%xmm6
1970	pxor %xmm7,%xmm7
1971	pxor %xmm2,%xmm1
1972	pxor %xmm0,%xmm2
1973	pxor %xmm3,%xmm1
1974	pxor %xmm0,%xmm3
1975	pxor %xmm4,%xmm1
1976	pxor %xmm0,%xmm4
1977	pxor %xmm5,%xmm1
1978	pxor %xmm0,%xmm5
1979	pxor %xmm6,%xmm1
1980	pxor %xmm0,%xmm6
1981	movdqa %xmm1,96(%esp)
1982	movups -32(%edx,%ecx,1),%xmm1
1983	pxor (%esp),%xmm2
1984	pxor 16(%esp),%xmm3
1985	pxor 32(%esp),%xmm4
1986	pxor 48(%esp),%xmm5
1987	pxor 64(%esp),%xmm6
1988	movups -16(%edx,%ecx,1),%xmm0
1989	.byte 102,15,56,220,209
1990	.byte 102,15,56,220,217
1991	.byte 102,15,56,220,225
1992	.byte 102,15,56,220,233
1993	.byte 102,15,56,220,241
1994	.byte 102,15,56,220,249
1995	movl 120(%esp),%edi
1996	call L_aesni_encrypt6_enter
1997	movdqa 64(%esp),%xmm0
1998	pxor (%esp),%xmm2
1999	pxor 16(%esp),%xmm3
2000	pxor 32(%esp),%xmm4
2001	pxor 48(%esp),%xmm5
2002	pxor %xmm0,%xmm6
2003	movdqa 96(%esp),%xmm1
2004	movdqu %xmm2,(%edi,%esi,1)
2005	movdqu %xmm3,16(%edi,%esi,1)
2006	movdqu %xmm4,32(%edi,%esi,1)
2007	movdqu %xmm5,48(%edi,%esi,1)
2008	movdqu %xmm6,64(%edi,%esi,1)
2009	jmp L078done
2010	.align 4,0x90
2011	L079one:
2012	movdqu (%ebx),%xmm7
2013	movl 112(%esp),%edx
2014	movdqu (%esi),%xmm2
2015	movl 240(%edx),%ecx
2016	pxor %xmm0,%xmm7
2017	pxor %xmm2,%xmm1
2018	pxor %xmm7,%xmm2
2019	movdqa %xmm1,%xmm6
2020	movl 120(%esp),%edi
2021	movups (%edx),%xmm0
2022	movups 16(%edx),%xmm1
2023	leal 32(%edx),%edx
2024	xorps %xmm0,%xmm2
2025	L083enc1_loop_16:
2026	.byte 102,15,56,220,209
2027	decl %ecx
2028	movups (%edx),%xmm1
2029	leal 16(%edx),%edx
2030	jnz L083enc1_loop_16
2031	.byte 102,15,56,221,209
2032	xorps %xmm7,%xmm2
2033	movdqa %xmm7,%xmm0
2034	movdqa %xmm6,%xmm1
2035	movups %xmm2,(%edi,%esi,1)
2036	jmp L078done
2037	.align 4,0x90
2038	L080two:
2039	leal 1(%ebp),%ecx
2040	movl 112(%esp),%edx
2041	bsfl %ecx,%ecx
2042	shll $4,%ecx
2043	movdqu (%ebx),%xmm6
2044	movdqu (%ebx,%ecx,1),%xmm7
2045	movdqu (%esi),%xmm2
2046	movdqu 16(%esi),%xmm3
2047	movl 240(%edx),%ecx
2048	pxor %xmm0,%xmm6
2049	pxor %xmm6,%xmm7
2050	pxor %xmm2,%xmm1
2051	pxor %xmm6,%xmm2
2052	pxor %xmm3,%xmm1
2053	pxor %xmm7,%xmm3
2054	movdqa %xmm1,%xmm5
2055	movl 120(%esp),%edi
2056	call __aesni_encrypt2
2057	xorps %xmm6,%xmm2
2058	xorps %xmm7,%xmm3
2059	movdqa %xmm7,%xmm0
2060	movdqa %xmm5,%xmm1
2061	movups %xmm2,(%edi,%esi,1)
2062	movups %xmm3,16(%edi,%esi,1)
2063	jmp L078done
2064	.align 4,0x90
2065	L081three:
2066	leal 1(%ebp),%ecx
2067	movl 112(%esp),%edx
2068	bsfl %ecx,%ecx
2069	shll $4,%ecx
2070	movdqu (%ebx),%xmm5
2071	movdqu (%ebx,%ecx,1),%xmm6
2072	movdqa %xmm5,%xmm7
2073	movdqu (%esi),%xmm2
2074	movdqu 16(%esi),%xmm3
2075	movdqu 32(%esi),%xmm4
2076	movl 240(%edx),%ecx
2077	pxor %xmm0,%xmm5
2078	pxor %xmm5,%xmm6
2079	pxor %xmm6,%xmm7
2080	pxor %xmm2,%xmm1
2081	pxor %xmm5,%xmm2
2082	pxor %xmm3,%xmm1
2083	pxor %xmm6,%xmm3
2084	pxor %xmm4,%xmm1
2085	pxor %xmm7,%xmm4
2086	movdqa %xmm1,96(%esp)
2087	movl 120(%esp),%edi
2088	call __aesni_encrypt3
2089	xorps %xmm5,%xmm2
2090	xorps %xmm6,%xmm3
2091	xorps %xmm7,%xmm4
2092	movdqa %xmm7,%xmm0
2093	movdqa 96(%esp),%xmm1
2094	movups %xmm2,(%edi,%esi,1)
2095	movups %xmm3,16(%edi,%esi,1)
2096	movups %xmm4,32(%edi,%esi,1)
2097	jmp L078done
2098	.align 4,0x90
2099	L082four:
2100	leal 1(%ebp),%ecx
2101	leal 3(%ebp),%eax
2102	bsfl %ecx,%ecx
2103	bsfl %eax,%eax
2104	movl 112(%esp),%edx
2105	shll $4,%ecx
2106	shll $4,%eax
2107	movdqu (%ebx),%xmm4
2108	movdqu (%ebx,%ecx,1),%xmm5
2109	movdqa %xmm4,%xmm6
2110	movdqu (%ebx,%eax,1),%xmm7
2111	pxor %xmm0,%xmm4
2112	movdqu (%esi),%xmm2
2113	pxor %xmm4,%xmm5
2114	movdqu 16(%esi),%xmm3
2115	pxor %xmm5,%xmm6
2116	movdqa %xmm4,(%esp)
2117	pxor %xmm6,%xmm7
2118	movdqa %xmm5,16(%esp)
2119	movdqu 32(%esi),%xmm4
2120	movdqu 48(%esi),%xmm5
2121	movl 240(%edx),%ecx
2122	pxor %xmm2,%xmm1
2123	pxor (%esp),%xmm2
2124	pxor %xmm3,%xmm1
2125	pxor 16(%esp),%xmm3
2126	pxor %xmm4,%xmm1
2127	pxor %xmm6,%xmm4
2128	pxor %xmm5,%xmm1
2129	pxor %xmm7,%xmm5
2130	movdqa %xmm1,96(%esp)
2131	movl 120(%esp),%edi
2132	call __aesni_encrypt4
2133	xorps (%esp),%xmm2
2134	xorps 16(%esp),%xmm3
2135	xorps %xmm6,%xmm4
2136	movups %xmm2,(%edi,%esi,1)
2137	xorps %xmm7,%xmm5
2138	movups %xmm3,16(%edi,%esi,1)
2139	movdqa %xmm7,%xmm0
2140	movups %xmm4,32(%edi,%esi,1)
2141	movdqa 96(%esp),%xmm1
2142	movups %xmm5,48(%edi,%esi,1)
2143	L078done:
2144	movl 128(%esp),%edx
2145	pxor %xmm2,%xmm2
2146	pxor %xmm3,%xmm3
2147	movdqa %xmm2,(%esp)
2148	pxor %xmm4,%xmm4
2149	movdqa %xmm2,16(%esp)
2150	pxor %xmm5,%xmm5
2151	movdqa %xmm2,32(%esp)
2152	pxor %xmm6,%xmm6
2153	movdqa %xmm2,48(%esp)
2154	pxor %xmm7,%xmm7
2155	movdqa %xmm2,64(%esp)
2156	movdqa %xmm2,80(%esp)
2157	movdqa %xmm2,96(%esp)
2158	leal (%edx),%esp
2159	movl 40(%esp),%ecx
2160	movl 48(%esp),%ebx
2161	movdqu %xmm0,(%ecx)
2162	pxor %xmm0,%xmm0
2163	movdqu %xmm1,(%ebx)
2164	pxor %xmm1,%xmm1
2165	popl %edi
2166	popl %esi
2167	popl %ebx
2168	popl %ebp
2169	ret
2170	.globl _aesni_ocb_decrypt
2171	.align 4
2172	_aesni_ocb_decrypt:
2173	L_aesni_ocb_decrypt_begin:
2174	.byte 243,15,30,251
2175	pushl %ebp
2176	pushl %ebx
2177	pushl %esi
2178	pushl %edi
2179	movl 40(%esp),%ecx
2180	movl 48(%esp),%ebx
2181	movl 20(%esp),%esi
2182	movl 24(%esp),%edi
2183	movl 28(%esp),%eax
2184	movl 32(%esp),%edx
2185	movdqu (%ecx),%xmm0
2186	movl 36(%esp),%ebp
2187	movdqu (%ebx),%xmm1
2188	movl 44(%esp),%ebx
2189	movl %esp,%ecx
2190	subl $132,%esp
2191	andl $-16,%esp
2192	subl %esi,%edi
2193	shll $4,%eax
2194	leal -96(%esi,%eax,1),%eax
2195	movl %edi,120(%esp)
2196	movl %eax,124(%esp)
2197	movl %ecx,128(%esp)
2198	movl 240(%edx),%ecx
2199	testl $1,%ebp
2200	jnz L084odd
2201	bsfl %ebp,%eax
2202	addl $1,%ebp
2203	shll $4,%eax
2204	movdqu (%ebx,%eax,1),%xmm7
2205	movl %edx,%eax
2206	movdqu (%esi),%xmm2
2207	leal 16(%esi),%esi
2208	pxor %xmm0,%xmm7
2209	pxor %xmm7,%xmm2
2210	movdqa %xmm1,%xmm6
2211	movups (%edx),%xmm0
2212	movups 16(%edx),%xmm1
2213	leal 32(%edx),%edx
2214	xorps %xmm0,%xmm2
2215	L085dec1_loop_17:
2216	.byte 102,15,56,222,209
2217	decl %ecx
2218	movups (%edx),%xmm1
2219	leal 16(%edx),%edx
2220	jnz L085dec1_loop_17
2221	.byte 102,15,56,223,209
2222	xorps %xmm7,%xmm2
2223	movaps %xmm6,%xmm1
2224	movdqa %xmm7,%xmm0
2225	xorps %xmm2,%xmm1
2226	movups %xmm2,-16(%edi,%esi,1)
2227	movl 240(%eax),%ecx
2228	movl %eax,%edx
2229	movl 124(%esp),%eax
2230	L084odd:
2231	shll $4,%ecx
2232	movl $16,%edi
2233	subl %ecx,%edi
2234	movl %edx,112(%esp)
2235	leal 32(%edx,%ecx,1),%edx
2236	movl %edi,116(%esp)
2237	cmpl %eax,%esi
2238	ja L086short
2239	jmp L087grandloop
2240	.align 5,0x90
2241	L087grandloop:
2242	leal 1(%ebp),%ecx
2243	leal 3(%ebp),%eax
2244	leal 5(%ebp),%edi
2245	addl $6,%ebp
2246	bsfl %ecx,%ecx
2247	bsfl %eax,%eax
2248	bsfl %edi,%edi
2249	shll $4,%ecx
2250	shll $4,%eax
2251	shll $4,%edi
2252	movdqu (%ebx),%xmm2
2253	movdqu (%ebx,%ecx,1),%xmm3
2254	movl 116(%esp),%ecx
2255	movdqa %xmm2,%xmm4
2256	movdqu (%ebx,%eax,1),%xmm5
2257	movdqa %xmm2,%xmm6
2258	movdqu (%ebx,%edi,1),%xmm7
2259	pxor %xmm0,%xmm2
2260	pxor %xmm2,%xmm3
2261	movdqa %xmm2,(%esp)
2262	pxor %xmm3,%xmm4
2263	movdqa %xmm3,16(%esp)
2264	pxor %xmm4,%xmm5
2265	movdqa %xmm4,32(%esp)
2266	pxor %xmm5,%xmm6
2267	movdqa %xmm5,48(%esp)
2268	pxor %xmm6,%xmm7
2269	movdqa %xmm6,64(%esp)
2270	movdqa %xmm7,80(%esp)
2271	movups -48(%edx,%ecx,1),%xmm0
2272	movdqu (%esi),%xmm2
2273	movdqu 16(%esi),%xmm3
2274	movdqu 32(%esi),%xmm4
2275	movdqu 48(%esi),%xmm5
2276	movdqu 64(%esi),%xmm6
2277	movdqu 80(%esi),%xmm7
2278	leal 96(%esi),%esi
2279	movdqa %xmm1,96(%esp)
2280	pxor %xmm0,%xmm2
2281	pxor %xmm0,%xmm3
2282	pxor %xmm0,%xmm4
2283	pxor %xmm0,%xmm5
2284	pxor %xmm0,%xmm6
2285	pxor %xmm0,%xmm7
2286	movups -32(%edx,%ecx,1),%xmm1
2287	pxor (%esp),%xmm2
2288	pxor 16(%esp),%xmm3
2289	pxor 32(%esp),%xmm4
2290	pxor 48(%esp),%xmm5
2291	pxor 64(%esp),%xmm6
2292	pxor 80(%esp),%xmm7
2293	movups -16(%edx,%ecx,1),%xmm0
2294	.byte 102,15,56,222,209
2295	.byte 102,15,56,222,217
2296	.byte 102,15,56,222,225
2297	.byte 102,15,56,222,233
2298	.byte 102,15,56,222,241
2299	.byte 102,15,56,222,249
2300	movl 120(%esp),%edi
2301	movl 124(%esp),%eax
2302	call L_aesni_decrypt6_enter
2303	movdqa 80(%esp),%xmm0
2304	pxor (%esp),%xmm2
2305	movdqa 96(%esp),%xmm1
2306	pxor 16(%esp),%xmm3
2307	pxor 32(%esp),%xmm4
2308	pxor 48(%esp),%xmm5
2309	pxor 64(%esp),%xmm6
2310	pxor %xmm0,%xmm7
2311	pxor %xmm2,%xmm1
2312	movdqu %xmm2,-96(%edi,%esi,1)
2313	pxor %xmm3,%xmm1
2314	movdqu %xmm3,-80(%edi,%esi,1)
2315	pxor %xmm4,%xmm1
2316	movdqu %xmm4,-64(%edi,%esi,1)
2317	pxor %xmm5,%xmm1
2318	movdqu %xmm5,-48(%edi,%esi,1)
2319	pxor %xmm6,%xmm1
2320	movdqu %xmm6,-32(%edi,%esi,1)
2321	pxor %xmm7,%xmm1
2322	movdqu %xmm7,-16(%edi,%esi,1)
2323	cmpl %eax,%esi
2324	jb L087grandloop
2325	L086short:
2326	addl $96,%eax
2327	subl %esi,%eax
2328	jz L088done
2329	cmpl $32,%eax
2330	jb L089one
2331	je L090two
2332	cmpl $64,%eax
2333	jb L091three
2334	je L092four
2335	leal 1(%ebp),%ecx
2336	leal 3(%ebp),%eax
2337	bsfl %ecx,%ecx
2338	bsfl %eax,%eax
2339	shll $4,%ecx
2340	shll $4,%eax
2341	movdqu (%ebx),%xmm2
2342	movdqu (%ebx,%ecx,1),%xmm3
2343	movl 116(%esp),%ecx
2344	movdqa %xmm2,%xmm4
2345	movdqu (%ebx,%eax,1),%xmm5
2346	movdqa %xmm2,%xmm6
2347	pxor %xmm0,%xmm2
2348	pxor %xmm2,%xmm3
2349	movdqa %xmm2,(%esp)
2350	pxor %xmm3,%xmm4
2351	movdqa %xmm3,16(%esp)
2352	pxor %xmm4,%xmm5
2353	movdqa %xmm4,32(%esp)
2354	pxor %xmm5,%xmm6
2355	movdqa %xmm5,48(%esp)
2356	pxor %xmm6,%xmm7
2357	movdqa %xmm6,64(%esp)
2358	movups -48(%edx,%ecx,1),%xmm0
2359	movdqu (%esi),%xmm2
2360	movdqu 16(%esi),%xmm3
2361	movdqu 32(%esi),%xmm4
2362	movdqu 48(%esi),%xmm5
2363	movdqu 64(%esi),%xmm6
2364	pxor %xmm7,%xmm7
2365	movdqa %xmm1,96(%esp)
2366	pxor %xmm0,%xmm2
2367	pxor %xmm0,%xmm3
2368	pxor %xmm0,%xmm4
2369	pxor %xmm0,%xmm5
2370	pxor %xmm0,%xmm6
2371	movups -32(%edx,%ecx,1),%xmm1
2372	pxor (%esp),%xmm2
2373	pxor 16(%esp),%xmm3
2374	pxor 32(%esp),%xmm4
2375	pxor 48(%esp),%xmm5
2376	pxor 64(%esp),%xmm6
2377	movups -16(%edx,%ecx,1),%xmm0
2378	.byte 102,15,56,222,209
2379	.byte 102,15,56,222,217
2380	.byte 102,15,56,222,225
2381	.byte 102,15,56,222,233
2382	.byte 102,15,56,222,241
2383	.byte 102,15,56,222,249
2384	movl 120(%esp),%edi
2385	call L_aesni_decrypt6_enter
2386	movdqa 64(%esp),%xmm0
2387	pxor (%esp),%xmm2
2388	movdqa 96(%esp),%xmm1
2389	pxor 16(%esp),%xmm3
2390	pxor 32(%esp),%xmm4
2391	pxor 48(%esp),%xmm5
2392	pxor %xmm0,%xmm6
2393	pxor %xmm2,%xmm1
2394	movdqu %xmm2,(%edi,%esi,1)
2395	pxor %xmm3,%xmm1
2396	movdqu %xmm3,16(%edi,%esi,1)
2397	pxor %xmm4,%xmm1
2398	movdqu %xmm4,32(%edi,%esi,1)
2399	pxor %xmm5,%xmm1
2400	movdqu %xmm5,48(%edi,%esi,1)
2401	pxor %xmm6,%xmm1
2402	movdqu %xmm6,64(%edi,%esi,1)
2403	jmp L088done
2404	.align 4,0x90
2405	L089one:
2406	movdqu (%ebx),%xmm7
2407	movl 112(%esp),%edx
2408	movdqu (%esi),%xmm2
2409	movl 240(%edx),%ecx
2410	pxor %xmm0,%xmm7
2411	pxor %xmm7,%xmm2
2412	movdqa %xmm1,%xmm6
2413	movl 120(%esp),%edi
2414	movups (%edx),%xmm0
2415	movups 16(%edx),%xmm1
2416	leal 32(%edx),%edx
2417	xorps %xmm0,%xmm2
2418	L093dec1_loop_18:
2419	.byte 102,15,56,222,209
2420	decl %ecx
2421	movups (%edx),%xmm1
2422	leal 16(%edx),%edx
2423	jnz L093dec1_loop_18
2424	.byte 102,15,56,223,209
2425	xorps %xmm7,%xmm2
2426	movaps %xmm6,%xmm1
2427	movdqa %xmm7,%xmm0
2428	xorps %xmm2,%xmm1
2429	movups %xmm2,(%edi,%esi,1)
2430	jmp L088done
2431	.align 4,0x90
2432	L090two:
2433	leal 1(%ebp),%ecx
2434	movl 112(%esp),%edx
2435	bsfl %ecx,%ecx
2436	shll $4,%ecx
2437	movdqu (%ebx),%xmm6
2438	movdqu (%ebx,%ecx,1),%xmm7
2439	movdqu (%esi),%xmm2
2440	movdqu 16(%esi),%xmm3
2441	movl 240(%edx),%ecx
2442	movdqa %xmm1,%xmm5
2443	pxor %xmm0,%xmm6
2444	pxor %xmm6,%xmm7
2445	pxor %xmm6,%xmm2
2446	pxor %xmm7,%xmm3
2447	movl 120(%esp),%edi
2448	call __aesni_decrypt2
2449	xorps %xmm6,%xmm2
2450	xorps %xmm7,%xmm3
2451	movdqa %xmm7,%xmm0
2452	xorps %xmm2,%xmm5
2453	movups %xmm2,(%edi,%esi,1)
2454	xorps %xmm3,%xmm5
2455	movups %xmm3,16(%edi,%esi,1)
2456	movaps %xmm5,%xmm1
2457	jmp L088done
2458	.align 4,0x90
2459	L091three:
2460	leal 1(%ebp),%ecx
2461	movl 112(%esp),%edx
2462	bsfl %ecx,%ecx
2463	shll $4,%ecx
2464	movdqu (%ebx),%xmm5
2465	movdqu (%ebx,%ecx,1),%xmm6
2466	movdqa %xmm5,%xmm7
2467	movdqu (%esi),%xmm2
2468	movdqu 16(%esi),%xmm3
2469	movdqu 32(%esi),%xmm4
2470	movl 240(%edx),%ecx
2471	movdqa %xmm1,96(%esp)
2472	pxor %xmm0,%xmm5
2473	pxor %xmm5,%xmm6
2474	pxor %xmm6,%xmm7
2475	pxor %xmm5,%xmm2
2476	pxor %xmm6,%xmm3
2477	pxor %xmm7,%xmm4
2478	movl 120(%esp),%edi
2479	call __aesni_decrypt3
2480	movdqa 96(%esp),%xmm1
2481	xorps %xmm5,%xmm2
2482	xorps %xmm6,%xmm3
2483	xorps %xmm7,%xmm4
2484	movups %xmm2,(%edi,%esi,1)
2485	pxor %xmm2,%xmm1
2486	movdqa %xmm7,%xmm0
2487	movups %xmm3,16(%edi,%esi,1)
2488	pxor %xmm3,%xmm1
2489	movups %xmm4,32(%edi,%esi,1)
2490	pxor %xmm4,%xmm1
2491	jmp L088done
2492	.align 4,0x90
2493	L092four:
2494	leal 1(%ebp),%ecx
2495	leal 3(%ebp),%eax
2496	bsfl %ecx,%ecx
2497	bsfl %eax,%eax
2498	movl 112(%esp),%edx
2499	shll $4,%ecx
2500	shll $4,%eax
2501	movdqu (%ebx),%xmm4
2502	movdqu (%ebx,%ecx,1),%xmm5
2503	movdqa %xmm4,%xmm6
2504	movdqu (%ebx,%eax,1),%xmm7
2505	pxor %xmm0,%xmm4
2506	movdqu (%esi),%xmm2
2507	pxor %xmm4,%xmm5
2508	movdqu 16(%esi),%xmm3
2509	pxor %xmm5,%xmm6
2510	movdqa %xmm4,(%esp)
2511	pxor %xmm6,%xmm7
2512	movdqa %xmm5,16(%esp)
2513	movdqu 32(%esi),%xmm4
2514	movdqu 48(%esi),%xmm5
2515	movl 240(%edx),%ecx
2516	movdqa %xmm1,96(%esp)
2517	pxor (%esp),%xmm2
2518	pxor 16(%esp),%xmm3
2519	pxor %xmm6,%xmm4
2520	pxor %xmm7,%xmm5
2521	movl 120(%esp),%edi
2522	call __aesni_decrypt4
2523	movdqa 96(%esp),%xmm1
2524	xorps (%esp),%xmm2
2525	xorps 16(%esp),%xmm3
2526	xorps %xmm6,%xmm4
2527	movups %xmm2,(%edi,%esi,1)
2528	pxor %xmm2,%xmm1
2529	xorps %xmm7,%xmm5
2530	movups %xmm3,16(%edi,%esi,1)
2531	pxor %xmm3,%xmm1
2532	movdqa %xmm7,%xmm0
2533	movups %xmm4,32(%edi,%esi,1)
2534	pxor %xmm4,%xmm1
2535	movups %xmm5,48(%edi,%esi,1)
2536	pxor %xmm5,%xmm1
2537	L088done:
2538	movl 128(%esp),%edx
2539	pxor %xmm2,%xmm2
2540	pxor %xmm3,%xmm3
2541	movdqa %xmm2,(%esp)
2542	pxor %xmm4,%xmm4
2543	movdqa %xmm2,16(%esp)
2544	pxor %xmm5,%xmm5
2545	movdqa %xmm2,32(%esp)
2546	pxor %xmm6,%xmm6
2547	movdqa %xmm2,48(%esp)
2548	pxor %xmm7,%xmm7
2549	movdqa %xmm2,64(%esp)
2550	movdqa %xmm2,80(%esp)
2551	movdqa %xmm2,96(%esp)
2552	leal (%edx),%esp
2553	movl 40(%esp),%ecx
2554	movl 48(%esp),%ebx
2555	movdqu %xmm0,(%ecx)
2556	pxor %xmm0,%xmm0
2557	movdqu %xmm1,(%ebx)
2558	pxor %xmm1,%xmm1
2559	popl %edi
2560	popl %esi
2561	popl %ebx
2562	popl %ebp
2563	ret
2564	.globl _aesni_cbc_encrypt
2565	.align 4
2566	_aesni_cbc_encrypt:
2567	L_aesni_cbc_encrypt_begin:
2568	.byte 243,15,30,251
2569	pushl %ebp
2570	pushl %ebx
2571	pushl %esi
2572	pushl %edi
2573	movl 20(%esp),%esi
2574	movl %esp,%ebx
2575	movl 24(%esp),%edi
2576	subl $24,%ebx
2577	movl 28(%esp),%eax
2578	andl $-16,%ebx
2579	movl 32(%esp),%edx
2580	movl 36(%esp),%ebp
2581	testl %eax,%eax
2582	jz L094cbc_abort
2583	cmpl $0,40(%esp)
2584	xchgl %esp,%ebx
2585	movups (%ebp),%xmm7
2586	movl 240(%edx),%ecx
2587	movl %edx,%ebp
2588	movl %ebx,16(%esp)
2589	movl %ecx,%ebx
2590	je L095cbc_decrypt
2591	movaps %xmm7,%xmm2
2592	cmpl $16,%eax
2593	jb L096cbc_enc_tail
2594	subl $16,%eax
2595	jmp L097cbc_enc_loop
2596	.align 4,0x90
2597	L097cbc_enc_loop:
2598	movups (%esi),%xmm7
2599	leal 16(%esi),%esi
2600	movups (%edx),%xmm0
2601	movups 16(%edx),%xmm1
2602	xorps %xmm0,%xmm7
2603	leal 32(%edx),%edx
2604	xorps %xmm7,%xmm2
2605	L098enc1_loop_19:
2606	.byte 102,15,56,220,209
2607	decl %ecx
2608	movups (%edx),%xmm1
2609	leal 16(%edx),%edx
2610	jnz L098enc1_loop_19
2611	.byte 102,15,56,221,209
2612	movl %ebx,%ecx
2613	movl %ebp,%edx
2614	movups %xmm2,(%edi)
2615	leal 16(%edi),%edi
2616	subl $16,%eax
2617	jnc L097cbc_enc_loop
2618	addl $16,%eax
2619	jnz L096cbc_enc_tail
2620	movaps %xmm2,%xmm7
2621	pxor %xmm2,%xmm2
2622	jmp L099cbc_ret
2623	L096cbc_enc_tail:
2624	movl %eax,%ecx
2625	.long 2767451785
2626	movl $16,%ecx
2627	subl %eax,%ecx
2628	xorl %eax,%eax
2629	.long 2868115081
2630	leal -16(%edi),%edi
2631	movl %ebx,%ecx
2632	movl %edi,%esi
2633	movl %ebp,%edx
2634	jmp L097cbc_enc_loop
2635	.align 4,0x90
2636	L095cbc_decrypt:
2637	cmpl $80,%eax
2638	jbe L100cbc_dec_tail
2639	movaps %xmm7,(%esp)
2640	subl $80,%eax
2641	jmp L101cbc_dec_loop6_enter
2642	.align 4,0x90
2643	L102cbc_dec_loop6:
2644	movaps %xmm0,(%esp)
2645	movups %xmm7,(%edi)
2646	leal 16(%edi),%edi
2647	L101cbc_dec_loop6_enter:
2648	movdqu (%esi),%xmm2
2649	movdqu 16(%esi),%xmm3
2650	movdqu 32(%esi),%xmm4
2651	movdqu 48(%esi),%xmm5
2652	movdqu 64(%esi),%xmm6
2653	movdqu 80(%esi),%xmm7
2654	call __aesni_decrypt6
2655	movups (%esi),%xmm1
2656	movups 16(%esi),%xmm0
2657	xorps (%esp),%xmm2
2658	xorps %xmm1,%xmm3
2659	movups 32(%esi),%xmm1
2660	xorps %xmm0,%xmm4
2661	movups 48(%esi),%xmm0
2662	xorps %xmm1,%xmm5
2663	movups 64(%esi),%xmm1
2664	xorps %xmm0,%xmm6
2665	movups 80(%esi),%xmm0
2666	xorps %xmm1,%xmm7
2667	movups %xmm2,(%edi)
2668	movups %xmm3,16(%edi)
2669	leal 96(%esi),%esi
2670	movups %xmm4,32(%edi)
2671	movl %ebx,%ecx
2672	movups %xmm5,48(%edi)
2673	movl %ebp,%edx
2674	movups %xmm6,64(%edi)
2675	leal 80(%edi),%edi
2676	subl $96,%eax
2677	ja L102cbc_dec_loop6
2678	movaps %xmm7,%xmm2
2679	movaps %xmm0,%xmm7
2680	addl $80,%eax
2681	jle L103cbc_dec_clear_tail_collected
2682	movups %xmm2,(%edi)
2683	leal 16(%edi),%edi
2684	L100cbc_dec_tail:
2685	movups (%esi),%xmm2
2686	movaps %xmm2,%xmm6
2687	cmpl $16,%eax
2688	jbe L104cbc_dec_one
2689	movups 16(%esi),%xmm3
2690	movaps %xmm3,%xmm5
2691	cmpl $32,%eax
2692	jbe L105cbc_dec_two
2693	movups 32(%esi),%xmm4
2694	cmpl $48,%eax
2695	jbe L106cbc_dec_three
2696	movups 48(%esi),%xmm5
2697	cmpl $64,%eax
2698	jbe L107cbc_dec_four
2699	movups 64(%esi),%xmm6
2700	movaps %xmm7,(%esp)
2701	movups (%esi),%xmm2
2702	xorps %xmm7,%xmm7
2703	call __aesni_decrypt6
2704	movups (%esi),%xmm1
2705	movups 16(%esi),%xmm0
2706	xorps (%esp),%xmm2
2707	xorps %xmm1,%xmm3
2708	movups 32(%esi),%xmm1
2709	xorps %xmm0,%xmm4
2710	movups 48(%esi),%xmm0
2711	xorps %xmm1,%xmm5
2712	movups 64(%esi),%xmm7
2713	xorps %xmm0,%xmm6
2714	movups %xmm2,(%edi)
2715	movups %xmm3,16(%edi)
2716	pxor %xmm3,%xmm3
2717	movups %xmm4,32(%edi)
2718	pxor %xmm4,%xmm4
2719	movups %xmm5,48(%edi)
2720	pxor %xmm5,%xmm5
2721	leal 64(%edi),%edi
2722	movaps %xmm6,%xmm2
2723	pxor %xmm6,%xmm6
2724	subl $80,%eax
2725	jmp L108cbc_dec_tail_collected
2726	.align 4,0x90
2727	L104cbc_dec_one:
2728	movups (%edx),%xmm0
2729	movups 16(%edx),%xmm1
2730	leal 32(%edx),%edx
2731	xorps %xmm0,%xmm2
2732	L109dec1_loop_20:
2733	.byte 102,15,56,222,209
2734	decl %ecx
2735	movups (%edx),%xmm1
2736	leal 16(%edx),%edx
2737	jnz L109dec1_loop_20
2738	.byte 102,15,56,223,209
2739	xorps %xmm7,%xmm2
2740	movaps %xmm6,%xmm7
2741	subl $16,%eax
2742	jmp L108cbc_dec_tail_collected
2743	.align 4,0x90
2744	L105cbc_dec_two:
2745	call __aesni_decrypt2
2746	xorps %xmm7,%xmm2
2747	xorps %xmm6,%xmm3
2748	movups %xmm2,(%edi)
2749	movaps %xmm3,%xmm2
2750	pxor %xmm3,%xmm3
2751	leal 16(%edi),%edi
2752	movaps %xmm5,%xmm7
2753	subl $32,%eax
2754	jmp L108cbc_dec_tail_collected
2755	.align 4,0x90
2756	L106cbc_dec_three:
2757	call __aesni_decrypt3
2758	xorps %xmm7,%xmm2
2759	xorps %xmm6,%xmm3
2760	xorps %xmm5,%xmm4
2761	movups %xmm2,(%edi)
2762	movaps %xmm4,%xmm2
2763	pxor %xmm4,%xmm4
2764	movups %xmm3,16(%edi)
2765	pxor %xmm3,%xmm3
2766	leal 32(%edi),%edi
2767	movups 32(%esi),%xmm7
2768	subl $48,%eax
2769	jmp L108cbc_dec_tail_collected
2770	.align 4,0x90
2771	L107cbc_dec_four:
2772	call __aesni_decrypt4
2773	movups 16(%esi),%xmm1
2774	movups 32(%esi),%xmm0
2775	xorps %xmm7,%xmm2
2776	movups 48(%esi),%xmm7
2777	xorps %xmm6,%xmm3
2778	movups %xmm2,(%edi)
2779	xorps %xmm1,%xmm4
2780	movups %xmm3,16(%edi)
2781	pxor %xmm3,%xmm3
2782	xorps %xmm0,%xmm5
2783	movups %xmm4,32(%edi)
2784	pxor %xmm4,%xmm4
2785	leal 48(%edi),%edi
2786	movaps %xmm5,%xmm2
2787	pxor %xmm5,%xmm5
2788	subl $64,%eax
2789	jmp L108cbc_dec_tail_collected
2790	.align 4,0x90
2791	L103cbc_dec_clear_tail_collected:
2792	pxor %xmm3,%xmm3
2793	pxor %xmm4,%xmm4
2794	pxor %xmm5,%xmm5
2795	pxor %xmm6,%xmm6
2796	L108cbc_dec_tail_collected:
2797	andl $15,%eax
2798	jnz L110cbc_dec_tail_partial
2799	movups %xmm2,(%edi)
2800	pxor %xmm0,%xmm0
2801	jmp L099cbc_ret
2802	.align 4,0x90
2803	L110cbc_dec_tail_partial:
2804	movaps %xmm2,(%esp)
2805	pxor %xmm0,%xmm0
2806	movl $16,%ecx
2807	movl %esp,%esi
2808	subl %eax,%ecx
2809	.long 2767451785
2810	movdqa %xmm2,(%esp)
2811	L099cbc_ret:
2812	movl 16(%esp),%esp
2813	movl 36(%esp),%ebp
2814	pxor %xmm2,%xmm2
2815	pxor %xmm1,%xmm1
2816	movups %xmm7,(%ebp)
2817	pxor %xmm7,%xmm7
2818	L094cbc_abort:
2819	popl %edi
2820	popl %esi
2821	popl %ebx
2822	popl %ebp
2823	ret
2824	.align 4
2825	__aesni_set_encrypt_key:
2826	.byte 243,15,30,251
2827	pushl %ebp
2828	pushl %ebx
2829	testl %eax,%eax
2830	jz L111bad_pointer
2831	testl %edx,%edx
2832	jz L111bad_pointer
2833	call L112pic
2834	L112pic:
2835	popl %ebx
2836	leal Lkey_const-L112pic(%ebx),%ebx
2837	movl L_OPENSSL_ia32cap_P$non_lazy_ptr-Lkey_const(%ebx),%ebp
2838	movups (%eax),%xmm0
2839	xorps %xmm4,%xmm4
2840	movl 4(%ebp),%ebp
2841	leal 16(%edx),%edx
2842	andl $268437504,%ebp
2843	cmpl $256,%ecx
2844	je L11314rounds
2845	cmpl $192,%ecx
2846	je L11412rounds
2847	cmpl $128,%ecx
2848	jne L115bad_keybits
2849	.align 4,0x90
2850	L11610rounds:
2851	cmpl $268435456,%ebp
2852	je L11710rounds_alt
2853	movl $9,%ecx
2854	movups %xmm0,-16(%edx)
2855	.byte 102,15,58,223,200,1
2856	call L118key_128_cold
2857	.byte 102,15,58,223,200,2
2858	call L119key_128
2859	.byte 102,15,58,223,200,4
2860	call L119key_128
2861	.byte 102,15,58,223,200,8
2862	call L119key_128
2863	.byte 102,15,58,223,200,16
2864	call L119key_128
2865	.byte 102,15,58,223,200,32
2866	call L119key_128
2867	.byte 102,15,58,223,200,64
2868	call L119key_128
2869	.byte 102,15,58,223,200,128
2870	call L119key_128
2871	.byte 102,15,58,223,200,27
2872	call L119key_128
2873	.byte 102,15,58,223,200,54
2874	call L119key_128
2875	movups %xmm0,(%edx)
2876	movl %ecx,80(%edx)
2877	jmp L120good_key
2878	.align 4,0x90
2879	L119key_128:
2880	movups %xmm0,(%edx)
2881	leal 16(%edx),%edx
2882	L118key_128_cold:
2883	shufps $16,%xmm0,%xmm4
2884	xorps %xmm4,%xmm0
2885	shufps $140,%xmm0,%xmm4
2886	xorps %xmm4,%xmm0
2887	shufps $255,%xmm1,%xmm1
2888	xorps %xmm1,%xmm0
2889	ret
2890	.align 4,0x90
2891	L11710rounds_alt:
2892	movdqa (%ebx),%xmm5
2893	movl $8,%ecx
2894	movdqa 32(%ebx),%xmm4
2895	movdqa %xmm0,%xmm2
2896	movdqu %xmm0,-16(%edx)
2897	L121loop_key128:
2898	.byte 102,15,56,0,197
2899	.byte 102,15,56,221,196
2900	pslld $1,%xmm4
2901	leal 16(%edx),%edx
2902	movdqa %xmm2,%xmm3
2903	pslldq $4,%xmm2
2904	pxor %xmm2,%xmm3
2905	pslldq $4,%xmm2
2906	pxor %xmm2,%xmm3
2907	pslldq $4,%xmm2
2908	pxor %xmm3,%xmm2
2909	pxor %xmm2,%xmm0
2910	movdqu %xmm0,-16(%edx)
2911	movdqa %xmm0,%xmm2
2912	decl %ecx
2913	jnz L121loop_key128
2914	movdqa 48(%ebx),%xmm4
2915	.byte 102,15,56,0,197
2916	.byte 102,15,56,221,196
2917	pslld $1,%xmm4
2918	movdqa %xmm2,%xmm3
2919	pslldq $4,%xmm2
2920	pxor %xmm2,%xmm3
2921	pslldq $4,%xmm2
2922	pxor %xmm2,%xmm3
2923	pslldq $4,%xmm2
2924	pxor %xmm3,%xmm2
2925	pxor %xmm2,%xmm0
2926	movdqu %xmm0,(%edx)
2927	movdqa %xmm0,%xmm2
2928	.byte 102,15,56,0,197
2929	.byte 102,15,56,221,196
2930	movdqa %xmm2,%xmm3
2931	pslldq $4,%xmm2
2932	pxor %xmm2,%xmm3
2933	pslldq $4,%xmm2
2934	pxor %xmm2,%xmm3
2935	pslldq $4,%xmm2
2936	pxor %xmm3,%xmm2
2937	pxor %xmm2,%xmm0
2938	movdqu %xmm0,16(%edx)
2939	movl $9,%ecx
2940	movl %ecx,96(%edx)
2941	jmp L120good_key
2942	.align 4,0x90
2943	L11412rounds:
2944	movq 16(%eax),%xmm2
2945	cmpl $268435456,%ebp
2946	je L12212rounds_alt
2947	movl $11,%ecx
2948	movups %xmm0,-16(%edx)
2949	.byte 102,15,58,223,202,1
2950	call L123key_192a_cold
2951	.byte 102,15,58,223,202,2
2952	call L124key_192b
2953	.byte 102,15,58,223,202,4
2954	call L125key_192a
2955	.byte 102,15,58,223,202,8
2956	call L124key_192b
2957	.byte 102,15,58,223,202,16
2958	call L125key_192a
2959	.byte 102,15,58,223,202,32
2960	call L124key_192b
2961	.byte 102,15,58,223,202,64
2962	call L125key_192a
2963	.byte 102,15,58,223,202,128
2964	call L124key_192b
2965	movups %xmm0,(%edx)
2966	movl %ecx,48(%edx)
2967	jmp L120good_key
2968	.align 4,0x90
2969	L125key_192a:
2970	movups %xmm0,(%edx)
2971	leal 16(%edx),%edx
2972	.align 4,0x90
2973	L123key_192a_cold:
2974	movaps %xmm2,%xmm5
2975	L126key_192b_warm:
2976	shufps $16,%xmm0,%xmm4
2977	movdqa %xmm2,%xmm3
2978	xorps %xmm4,%xmm0
2979	shufps $140,%xmm0,%xmm4
2980	pslldq $4,%xmm3
2981	xorps %xmm4,%xmm0
2982	pshufd $85,%xmm1,%xmm1
2983	pxor %xmm3,%xmm2
2984	pxor %xmm1,%xmm0
2985	pshufd $255,%xmm0,%xmm3
2986	pxor %xmm3,%xmm2
2987	ret
2988	.align 4,0x90
2989	L124key_192b:
2990	movaps %xmm0,%xmm3
2991	shufps $68,%xmm0,%xmm5
2992	movups %xmm5,(%edx)
2993	shufps $78,%xmm2,%xmm3
2994	movups %xmm3,16(%edx)
2995	leal 32(%edx),%edx
2996	jmp L126key_192b_warm
2997	.align 4,0x90
2998	L12212rounds_alt:
2999	movdqa 16(%ebx),%xmm5
3000	movdqa 32(%ebx),%xmm4
3001	movl $8,%ecx
3002	movdqu %xmm0,-16(%edx)
3003	L127loop_key192:
3004	movq %xmm2,(%edx)
3005	movdqa %xmm2,%xmm1
3006	.byte 102,15,56,0,213
3007	.byte 102,15,56,221,212
3008	pslld $1,%xmm4
3009	leal 24(%edx),%edx
3010	movdqa %xmm0,%xmm3
3011	pslldq $4,%xmm0
3012	pxor %xmm0,%xmm3
3013	pslldq $4,%xmm0
3014	pxor %xmm0,%xmm3
3015	pslldq $4,%xmm0
3016	pxor %xmm3,%xmm0
3017	pshufd $255,%xmm0,%xmm3
3018	pxor %xmm1,%xmm3
3019	pslldq $4,%xmm1
3020	pxor %xmm1,%xmm3
3021	pxor %xmm2,%xmm0
3022	pxor %xmm3,%xmm2
3023	movdqu %xmm0,-16(%edx)
3024	decl %ecx
3025	jnz L127loop_key192
3026	movl $11,%ecx
3027	movl %ecx,32(%edx)
3028	jmp L120good_key
3029	.align 4,0x90
3030	L11314rounds:
3031	movups 16(%eax),%xmm2
3032	leal 16(%edx),%edx
3033	cmpl $268435456,%ebp
3034	je L12814rounds_alt
3035	movl $13,%ecx
3036	movups %xmm0,-32(%edx)
3037	movups %xmm2,-16(%edx)
3038	.byte 102,15,58,223,202,1
3039	call L129key_256a_cold
3040	.byte 102,15,58,223,200,1
3041	call L130key_256b
3042	.byte 102,15,58,223,202,2
3043	call L131key_256a
3044	.byte 102,15,58,223,200,2
3045	call L130key_256b
3046	.byte 102,15,58,223,202,4
3047	call L131key_256a
3048	.byte 102,15,58,223,200,4
3049	call L130key_256b
3050	.byte 102,15,58,223,202,8
3051	call L131key_256a
3052	.byte 102,15,58,223,200,8
3053	call L130key_256b
3054	.byte 102,15,58,223,202,16
3055	call L131key_256a
3056	.byte 102,15,58,223,200,16
3057	call L130key_256b
3058	.byte 102,15,58,223,202,32
3059	call L131key_256a
3060	.byte 102,15,58,223,200,32
3061	call L130key_256b
3062	.byte 102,15,58,223,202,64
3063	call L131key_256a
3064	movups %xmm0,(%edx)
3065	movl %ecx,16(%edx)
3066	xorl %eax,%eax
3067	jmp L120good_key
3068	.align 4,0x90
3069	L131key_256a:
3070	movups %xmm2,(%edx)
3071	leal 16(%edx),%edx
3072	L129key_256a_cold:
3073	shufps $16,%xmm0,%xmm4
3074	xorps %xmm4,%xmm0
3075	shufps $140,%xmm0,%xmm4
3076	xorps %xmm4,%xmm0
3077	shufps $255,%xmm1,%xmm1
3078	xorps %xmm1,%xmm0
3079	ret
3080	.align 4,0x90
3081	L130key_256b:
3082	movups %xmm0,(%edx)
3083	leal 16(%edx),%edx
3084	shufps $16,%xmm2,%xmm4
3085	xorps %xmm4,%xmm2
3086	shufps $140,%xmm2,%xmm4
3087	xorps %xmm4,%xmm2
3088	shufps $170,%xmm1,%xmm1
3089	xorps %xmm1,%xmm2
3090	ret
3091	.align 4,0x90
3092	L12814rounds_alt:
3093	movdqa (%ebx),%xmm5
3094	movdqa 32(%ebx),%xmm4
3095	movl $7,%ecx
3096	movdqu %xmm0,-32(%edx)
3097	movdqa %xmm2,%xmm1
3098	movdqu %xmm2,-16(%edx)
3099	L132loop_key256:
3100	.byte 102,15,56,0,213
3101	.byte 102,15,56,221,212
3102	movdqa %xmm0,%xmm3
3103	pslldq $4,%xmm0
3104	pxor %xmm0,%xmm3
3105	pslldq $4,%xmm0
3106	pxor %xmm0,%xmm3
3107	pslldq $4,%xmm0
3108	pxor %xmm3,%xmm0
3109	pslld $1,%xmm4
3110	pxor %xmm2,%xmm0
3111	movdqu %xmm0,(%edx)
3112	decl %ecx
3113	jz L133done_key256
3114	pshufd $255,%xmm0,%xmm2
3115	pxor %xmm3,%xmm3
3116	.byte 102,15,56,221,211
3117	movdqa %xmm1,%xmm3
3118	pslldq $4,%xmm1
3119	pxor %xmm1,%xmm3
3120	pslldq $4,%xmm1
3121	pxor %xmm1,%xmm3
3122	pslldq $4,%xmm1
3123	pxor %xmm3,%xmm1
3124	pxor %xmm1,%xmm2
3125	movdqu %xmm2,16(%edx)
3126	leal 32(%edx),%edx
3127	movdqa %xmm2,%xmm1
3128	jmp L132loop_key256
3129	L133done_key256:
3130	movl $13,%ecx
3131	movl %ecx,16(%edx)
3132	L120good_key:
3133	pxor %xmm0,%xmm0
3134	pxor %xmm1,%xmm1
3135	pxor %xmm2,%xmm2
3136	pxor %xmm3,%xmm3
3137	pxor %xmm4,%xmm4
3138	pxor %xmm5,%xmm5
3139	xorl %eax,%eax
3140	popl %ebx
3141	popl %ebp
3142	ret
3143	.align 2,0x90
3144	L111bad_pointer:
3145	movl $-1,%eax
3146	popl %ebx
3147	popl %ebp
3148	ret
3149	.align 2,0x90
3150	L115bad_keybits:
3151	pxor %xmm0,%xmm0
3152	movl $-2,%eax
3153	popl %ebx
3154	popl %ebp
3155	ret
3156	.globl _aesni_set_encrypt_key
3157	.align 4
3158	_aesni_set_encrypt_key:
3159	L_aesni_set_encrypt_key_begin:
3160	.byte 243,15,30,251
3161	movl 4(%esp),%eax
3162	movl 8(%esp),%ecx
3163	movl 12(%esp),%edx
3164	call __aesni_set_encrypt_key
3165	ret
3166	.globl _aesni_set_decrypt_key
3167	.align 4
3168	_aesni_set_decrypt_key:
3169	L_aesni_set_decrypt_key_begin:
3170	.byte 243,15,30,251
3171	movl 4(%esp),%eax
3172	movl 8(%esp),%ecx
3173	movl 12(%esp),%edx
3174	call __aesni_set_encrypt_key
3175	movl 12(%esp),%edx
3176	shll $4,%ecx
3177	testl %eax,%eax
3178	jnz L134dec_key_ret
3179	leal 16(%edx,%ecx,1),%eax
3180	movups (%edx),%xmm0
3181	movups (%eax),%xmm1
3182	movups %xmm0,(%eax)
3183	movups %xmm1,(%edx)
3184	leal 16(%edx),%edx
3185	leal -16(%eax),%eax
3186	L135dec_key_inverse:
3187	movups (%edx),%xmm0
3188	movups (%eax),%xmm1
3189	.byte 102,15,56,219,192
3190	.byte 102,15,56,219,201
3191	leal 16(%edx),%edx
3192	leal -16(%eax),%eax
3193	movups %xmm0,16(%eax)
3194	movups %xmm1,-16(%edx)
3195	cmpl %edx,%eax
3196	ja L135dec_key_inverse
3197	movups (%edx),%xmm0
3198	.byte 102,15,56,219,192
3199	movups %xmm0,(%edx)
3200	pxor %xmm0,%xmm0
3201	pxor %xmm1,%xmm1
3202	xorl %eax,%eax
3203	L134dec_key_ret:
3204	ret
3205	.align 6,0x90
3206	Lkey_const:
3207	.long 202313229,202313229,202313229,202313229
3208	.long 67569157,67569157,67569157,67569157
3209	.long 1,1,1,1
3210	.long 27,27,27,27
3211	.byte 65,69,83,32,102,111,114,32,73,110,116,101,108,32,65,69
3212	.byte 83,45,78,73,44,32,67,82,89,80,84,79,71,65,77,83
3213	.byte 32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115
3214	.byte 115,108,46,111,114,103,62,0
3215	.section __IMPORT,__pointers,non_lazy_symbol_pointers
3216	L_OPENSSL_ia32cap_P$non_lazy_ptr:
3217	.indirect_symbol _OPENSSL_ia32cap_P
3218	.long 0
3219	.comm _OPENSSL_ia32cap_P,16,2

Note: See TracBrowser for help on using the repository browser.

Download in other formats:

Original Format