xref: /freebsd/sys/crypto/openssl/powerpc64le/vpaes-ppc.S (revision 02e9120893770924227138ba49df1edb3896112a)
1/* Do not modify. This file is auto-generated from vpaes-ppc.pl. */
2.machine	"any"
3
4.abiversion	2
5.text
6
7.align	7
8_vpaes_consts:
9.Lk_mc_forward:
10.byte	0x0e,0x0d,0x0c,0x0f,0x0a,0x09,0x08,0x0b,0x06,0x05,0x04,0x07,0x02,0x01,0x00,0x03
11.byte	0x0a,0x09,0x08,0x0b,0x06,0x05,0x04,0x07,0x02,0x01,0x00,0x03,0x0e,0x0d,0x0c,0x0f
12.byte	0x06,0x05,0x04,0x07,0x02,0x01,0x00,0x03,0x0e,0x0d,0x0c,0x0f,0x0a,0x09,0x08,0x0b
13.byte	0x02,0x01,0x00,0x03,0x0e,0x0d,0x0c,0x0f,0x0a,0x09,0x08,0x0b,0x06,0x05,0x04,0x07
14.Lk_mc_backward:
15.byte	0x0c,0x0f,0x0e,0x0d,0x08,0x0b,0x0a,0x09,0x04,0x07,0x06,0x05,0x00,0x03,0x02,0x01
16.byte	0x00,0x03,0x02,0x01,0x0c,0x0f,0x0e,0x0d,0x08,0x0b,0x0a,0x09,0x04,0x07,0x06,0x05
17.byte	0x04,0x07,0x06,0x05,0x00,0x03,0x02,0x01,0x0c,0x0f,0x0e,0x0d,0x08,0x0b,0x0a,0x09
18.byte	0x08,0x0b,0x0a,0x09,0x04,0x07,0x06,0x05,0x00,0x03,0x02,0x01,0x0c,0x0f,0x0e,0x0d
19.Lk_sr:
20.byte	0x0f,0x0e,0x0d,0x0c,0x0b,0x0a,0x09,0x08,0x07,0x06,0x05,0x04,0x03,0x02,0x01,0x00
21.byte	0x0f,0x0a,0x05,0x00,0x0b,0x06,0x01,0x0c,0x07,0x02,0x0d,0x08,0x03,0x0e,0x09,0x04
22.byte	0x0f,0x06,0x0d,0x04,0x0b,0x02,0x09,0x00,0x07,0x0e,0x05,0x0c,0x03,0x0a,0x01,0x08
23.byte	0x0f,0x02,0x05,0x08,0x0b,0x0e,0x01,0x04,0x07,0x0a,0x0d,0x00,0x03,0x06,0x09,0x0c
24
25
26
27
28.Lk_inv:
29.byte	0x04,0x07,0x03,0x09,0x0a,0x0b,0x0c,0x02,0x0e,0x05,0x06,0x0f,0x0d,0x08,0x01,0xf0
30.byte	0x03,0x0d,0x0e,0x0c,0x02,0x05,0x08,0x09,0x01,0x04,0x0a,0x06,0x0f,0x0b,0x07,0xf0
31.Lk_ipt:
32.byte	0xca,0xba,0xe0,0x90,0x52,0x22,0x78,0x08,0xc2,0xb2,0xe8,0x98,0x5a,0x2a,0x70,0x00
33.byte	0xcd,0x80,0xb1,0xfc,0xb0,0xfd,0xcc,0x81,0x4c,0x01,0x30,0x7d,0x31,0x7c,0x4d,0x00
34.Lk_sbo:
35.byte	0x15,0xaa,0xbf,0x7a,0xc5,0x02,0xa8,0x78,0xd0,0xd2,0x6d,0x17,0x6f,0xbd,0xc7,0x00
36.byte	0x8e,0x1e,0x90,0xd1,0x41,0x2b,0x35,0xfa,0xcf,0xe4,0x74,0xa5,0x5f,0xbb,0x6a,0x00
37.Lk_sb1:
38.byte	0x3b,0xf7,0xcc,0xc1,0x0d,0x2e,0xd9,0xef,0x36,0x18,0xd4,0x15,0xfa,0xe2,0x23,0x00
39.byte	0xa5,0xdf,0x7a,0x6e,0x14,0x2a,0xf5,0x44,0xb1,0x9b,0xe1,0x8f,0xcb,0x50,0x3e,0x00
40.Lk_sb2:
41.byte	0xc2,0xa1,0x63,0xc8,0xab,0x82,0x23,0x4a,0x69,0xeb,0x88,0x40,0x0a,0xe1,0x29,0x00
42.byte	0x5e,0xb7,0xe9,0x55,0xbc,0x98,0x2f,0xcd,0xe2,0x7a,0x93,0xc6,0x0b,0x71,0x24,0x00
43
44
45
46
47.Lk_dipt:
48.byte	0x15,0x4a,0x41,0x1e,0x11,0x4e,0x45,0x1a,0x0f,0x50,0x5b,0x04,0x0b,0x54,0x5f,0x00
49.byte	0x12,0x77,0x17,0x72,0xf4,0x91,0xf1,0x94,0x86,0xe3,0x83,0xe6,0x60,0x05,0x65,0x00
50.Lk_dsbo:
51.byte	0xc7,0xaa,0x6d,0xb9,0xd4,0x94,0x3e,0x2d,0x13,0x87,0xea,0x53,0x7e,0xf9,0x40,0x00
52.byte	0xca,0x4b,0x81,0x59,0xd8,0xc5,0x8e,0x9c,0x12,0xd7,0x56,0x0f,0x93,0x44,0x1d,0x00
53.Lk_dsb9:
54.byte	0xca,0xd5,0x1f,0x50,0x4f,0x99,0x4c,0xc9,0x85,0x1c,0x03,0x53,0x9a,0x86,0xd6,0x00
55.byte	0x72,0x5e,0x2c,0x9e,0xb2,0xfb,0xa5,0x65,0xc0,0x3b,0x17,0x89,0xec,0xd7,0x49,0x00
56.Lk_dsbd:
57.byte	0xf5,0x6e,0x9b,0x13,0x88,0x2a,0x44,0x39,0x7d,0x57,0xcc,0xdf,0xe6,0xb1,0xa2,0x00
58.byte	0x29,0x31,0x18,0x0d,0x15,0xde,0xef,0xd3,0x3c,0xe2,0xfa,0xf7,0x24,0xc6,0xcb,0x00
59.Lk_dsbb:
60.byte	0x60,0x26,0x46,0xf6,0xb0,0xf2,0xd4,0x04,0xd0,0x22,0x64,0x92,0x96,0xb4,0x42,0x00
61.byte	0xf3,0xff,0x0c,0x3e,0x32,0x55,0xaa,0x6b,0xc1,0x94,0x98,0xa6,0xcd,0x59,0x67,0x00
62.Lk_dsbe:
63.byte	0x22,0x42,0x60,0x04,0x64,0xb4,0xf6,0xb0,0x46,0xf2,0x92,0x96,0x26,0xd4,0xd0,0x00
64.byte	0x94,0x67,0xf3,0x6b,0x98,0x59,0x3e,0x32,0x0c,0x55,0xa6,0xcd,0xff,0xaa,0xc1,0x00
65
66
67
68
69.Lk_dksd:
70.byte	0x07,0x40,0xe3,0xa4,0x5a,0x1d,0xbe,0xf9,0xfe,0xb9,0x1a,0x5d,0xa3,0xe4,0x47,0x00
71.byte	0x5f,0xdc,0x69,0xea,0xab,0x28,0x9d,0x1e,0x41,0xc2,0x77,0xf4,0xb5,0x36,0x83,0x00
72.Lk_dksb:
73.byte	0x03,0xd6,0x53,0x86,0x1c,0xc9,0x4c,0x99,0x9a,0x4f,0xca,0x1f,0x85,0x50,0xd5,0x00
74.byte	0xd9,0x93,0x25,0x6f,0x7e,0x34,0x82,0xc8,0x11,0x5b,0xed,0xa7,0xb6,0xfc,0x4a,0x00
75.Lk_dkse:
76.byte	0x53,0x85,0x9a,0x4c,0x99,0x4f,0x50,0x86,0xd5,0x03,0x1c,0xca,0x1f,0xc9,0xd6,0x00
77.byte	0xcd,0x5e,0xf9,0x6a,0x20,0xb3,0x14,0x87,0xa2,0x31,0x96,0x05,0x4f,0xdc,0x7b,0xe8
78.Lk_dks9:
79.byte	0x4a,0xed,0x93,0x34,0x82,0x25,0x5b,0xfc,0xb6,0x11,0x6f,0xc8,0x7e,0xd9,0xa7,0x00
80.byte	0x8b,0xb8,0x9f,0xac,0xe9,0xda,0xfd,0xce,0x45,0x76,0x51,0x62,0x27,0x14,0x33,0x00
81
82.Lk_rcon:
83.byte	0xb6,0xee,0x9d,0xaf,0xb9,0x91,0x83,0x1f,0x81,0x7d,0x7c,0x4d,0x08,0x98,0x2a,0x70
84.Lk_s63:
85.byte	0x5b,0x5b,0x5b,0x5b,0x5b,0x5b,0x5b,0x5b,0x5b,0x5b,0x5b,0x5b,0x5b,0x5b,0x5b,0x5b
86
87.Lk_opt:
88.byte	0xf7,0x97,0x41,0x21,0xde,0xbe,0x68,0x08,0xff,0x9f,0x49,0x29,0xd6,0xb6,0x60,0x00
89.byte	0xe1,0x0d,0x5d,0xb1,0xb0,0x5c,0x0c,0xe0,0x01,0xed,0xbd,0x51,0x50,0xbc,0xec,0x00
90.Lk_deskew:
91.byte	0x1d,0xfe,0xb9,0x5a,0x5d,0xbe,0xf9,0x1a,0x07,0xe4,0xa3,0x40,0x47,0xa4,0xe3,0x00
92.byte	0x28,0x41,0xc2,0xab,0xf4,0x9d,0x1e,0x77,0x5f,0x36,0xb5,0xdc,0x83,0xea,0x69,0x00
93.align	5
94.Lconsts:
95	mflr	0
96	bcl	20,31,$+4
97	mflr	12
98	addi	12,12,-0x308
99	mtlr	0
100	blr
101.long	0
102.byte	0,12,0x14,0,0,0,0,0
103.byte	86,101,99,116,111,114,32,80,101,114,109,117,116,97,116,105,111,110,32,65,69,83,32,102,111,114,32,65,108,116,105,86,101,99,44,77,105,107,101,32,72,97,109,98,117,114,103,32,40,83,116,97,110,102,111,114,100,32,85,110,105,118,101,114,115,105,116,121,41,0
104.align	2
105.align	6
106
107
108
109
110
111
112.align	4
113_vpaes_encrypt_preheat:
114	mflr	8
115	bl	.Lconsts
116	mtlr	8
117	li	11, 0xc0
118	li	10, 0xd0
119	li	9,  0xe0
120	li	8,  0xf0
121	vxor	7, 7, 7
122	vspltisb	8,4
123	vspltisb	9,0x0f
124	lvx	10, 12, 11
125	li	11, 0x100
126	lvx	11, 12, 10
127	li	10, 0x110
128	lvx	12, 12, 9
129	li	9,  0x120
130	lvx	13, 12, 8
131	li	8,  0x130
132	lvx	14, 12, 11
133	li	11, 0x140
134	lvx	15, 12, 10
135	li	10, 0x150
136	lvx	16, 12, 9
137	lvx	17, 12, 8
138	lvx	18, 12, 11
139	lvx	19, 12, 10
140	blr
141.long	0
142.byte	0,12,0x14,0,0,0,0,0
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158.align	5
159_vpaes_encrypt_core:
160	lwz	8, 240(5)
161	li	9, 16
162	lvx	5, 0, 5
163	li	11, 0x10
164	lvx	6, 9, 5
165	addi	9, 9, 16
166	vperm	5, 6, 5, 31
167	addi	10, 11, 0x40
168	vsrb	1, 0, 8
169	vperm	0, 12, 12, 0
170	vperm	1, 13, 13, 1
171	vxor	0, 0, 5
172	vxor	0, 0, 1
173	mtctr	8
174	b	.Lenc_entry
175
176.align	4
177.Lenc_loop:
178
179	vperm	4, 17, 7, 2
180	lvx	1, 12, 11
181	addi	11, 11, 16
182	vperm	0, 16, 7, 3
183	vxor	4, 4, 5
184	andi.	11, 11, 0x30
185	vperm	5, 19, 7, 2
186	vxor	0, 0, 4
187	vperm	2, 18, 7, 3
188	lvx	4, 12, 10
189	addi	10, 11, 0x40
190	vperm	3, 0, 7, 1
191	vxor	2, 2, 5
192	vperm	0, 0, 7, 4
193	vxor	3, 3, 2
194	vperm	4, 3, 7, 1
195	vxor	0, 0, 3
196	vxor	0, 0, 4
197
198.Lenc_entry:
199
200	vsrb	1, 0, 8
201	vperm	5, 11, 11, 0
202	vxor	0, 0, 1
203	vperm	3, 10, 10, 1
204	vperm	4, 10, 10, 0
205	vand	0, 0, 9
206	vxor	3, 3, 5
207	vxor	4, 4, 5
208	vperm	2, 10, 7, 3
209	vor	5,6,6
210	lvx	6, 9, 5
211	vperm	3, 10, 7, 4
212	addi	9, 9, 16
213	vxor	2, 2, 0
214	vperm	5, 6, 5, 31
215	vxor	3, 3, 1
216	bdnz	.Lenc_loop
217
218
219	addi	10, 11, 0x80
220
221
222	vperm	4, 14, 7, 2
223	lvx	1, 12, 10
224	vperm	0, 15, 7, 3
225	vxor	4, 4, 5
226	vxor	0, 0, 4
227	vperm	0, 0, 7, 1
228	blr
229.long	0
230.byte	0,12,0x14,0,0,0,0,0
231
232.globl	vpaes_encrypt
233.type	vpaes_encrypt,@function
234.align	5
235vpaes_encrypt:
236.localentry	vpaes_encrypt,0
237
238	stdu	1,-256(1)
239	li	10,63
240	li	11,79
241	mflr	6
242	li	7,-1
243	stvx	20,10,1
244	addi	10,10,32
245	stvx	21,11,1
246	addi	11,11,32
247	stvx	22,10,1
248	addi	10,10,32
249	stvx	23,11,1
250	addi	11,11,32
251	stvx	24,10,1
252	addi	10,10,32
253	stvx	25,11,1
254	addi	11,11,32
255	stvx	26,10,1
256	addi	10,10,32
257	stvx	27,11,1
258	addi	11,11,32
259	stvx	28,10,1
260	addi	10,10,32
261	stvx	29,11,1
262	addi	11,11,32
263	stvx	30,10,1
264	stvx	31,11,1
265	stw	7,252(1)
266	li	0, -1
267	std	6,272(1)
268	or	0,0,0
269
270	bl	_vpaes_encrypt_preheat
271
272	lvsr	27, 0, 3
273	lvx	0, 0, 3
274	addi	3, 3, 15
275	lvsl	29, 0, 4
276	lvsr	31, 0, 5
277	lvx	26, 0, 3
278	vperm	0, 26, 0, 27
279
280	bl	_vpaes_encrypt_core
281
282	andi.	8, 4, 15
283	li	9, 16
284	beq	.Lenc_out_aligned
285
286	vperm	0, 0, 0, 29
287	mtctr	9
288.Lenc_out_unaligned:
289	stvebx	0, 0, 4
290	addi	4, 4, 1
291	bdnz	.Lenc_out_unaligned
292	b	.Lenc_done
293
294.align	4
295.Lenc_out_aligned:
296	stvx	0, 0, 4
297.Lenc_done:
298
299	li	10,63
300	li	11,79
301	mtlr	6
302	or	7,7,7
303	lvx	20,10,1
304	addi	10,10,32
305	lvx	21,11,1
306	addi	11,11,32
307	lvx	22,10,1
308	addi	10,10,32
309	lvx	23,11,1
310	addi	11,11,32
311	lvx	24,10,1
312	addi	10,10,32
313	lvx	25,11,1
314	addi	11,11,32
315	lvx	26,10,1
316	addi	10,10,32
317	lvx	27,11,1
318	addi	11,11,32
319	lvx	28,10,1
320	addi	10,10,32
321	lvx	29,11,1
322	addi	11,11,32
323	lvx	30,10,1
324	lvx	31,11,1
325	addi	1,1,256
326	blr
327.long	0
328.byte	0,12,0x04,1,0x80,0,3,0
329.long	0
330.size	vpaes_encrypt,.-vpaes_encrypt
331
332.align	4
333_vpaes_decrypt_preheat:
334	mflr	8
335	bl	.Lconsts
336	mtlr	8
337	li	11, 0xc0
338	li	10, 0xd0
339	li	9,  0x160
340	li	8,  0x170
341	vxor	7, 7, 7
342	vspltisb	8,4
343	vspltisb	9,0x0f
344	lvx	10, 12, 11
345	li	11, 0x180
346	lvx	11, 12, 10
347	li	10, 0x190
348	lvx	12, 12, 9
349	li	9,  0x1a0
350	lvx	13, 12, 8
351	li	8,  0x1b0
352	lvx	14, 12, 11
353	li	11, 0x1c0
354	lvx	15, 12, 10
355	li	10, 0x1d0
356	lvx	16, 12, 9
357	li	9,  0x1e0
358	lvx	17, 12, 8
359	li	8,  0x1f0
360	lvx	18, 12, 11
361	li	11, 0x200
362	lvx	19, 12, 10
363	li	10, 0x210
364	lvx	20, 12, 9
365	lvx	21, 12, 8
366	lvx	22, 12, 11
367	lvx	23, 12, 10
368	blr
369.long	0
370.byte	0,12,0x14,0,0,0,0,0
371
372
373
374
375
376
377.align	4
378_vpaes_decrypt_core:
379	lwz	8, 240(5)
380	li	9, 16
381	lvx	5, 0, 5
382	li	11, 0x30
383	lvx	6, 9, 5
384	addi	9, 9, 16
385	vperm	5, 6, 5, 31
386	vsrb	1, 0, 8
387	vperm	0, 12, 12, 0
388	vperm	1, 13, 13, 1
389	vxor	0, 0, 5
390	vxor	0, 0, 1
391	mtctr	8
392	b	.Ldec_entry
393
394.align	4
395.Ldec_loop:
396
397
398
399	lvx	0, 12, 11
400
401
402	vperm	4, 16, 7, 2
403	subi	11, 11, 16
404	vperm	1, 17, 7, 3
405	andi.	11, 11, 0x30
406	vxor	5, 5, 4
407
408	vxor	5, 5, 1
409
410
411	vperm	4, 18, 7, 2
412	vperm	5, 5, 7, 0
413	vperm	1, 19, 7, 3
414	vxor	5, 5, 4
415
416	vxor	5, 5, 1
417
418
419	vperm	4, 20, 7, 2
420	vperm	5, 5, 7, 0
421	vperm	1, 21, 7, 3
422	vxor	5, 5, 4
423
424	vxor	5, 5, 1
425
426
427	vperm	4, 22, 7, 2
428	vperm	5, 5, 7, 0
429	vperm	1, 23, 7, 3
430	vxor	0, 5, 4
431	vxor	0, 0, 1
432
433.Ldec_entry:
434
435	vsrb	1, 0, 8
436	vperm	2, 11, 11, 0
437	vxor	0, 0, 1
438	vperm	3, 10, 10, 1
439	vperm	4, 10, 10, 0
440	vand	0, 0, 9
441	vxor	3, 3, 2
442	vxor	4, 4, 2
443	vperm	2, 10, 7, 3
444	vor	5,6,6
445	lvx	6, 9, 5
446	vperm	3, 10, 7, 4
447	addi	9, 9, 16
448	vxor	2, 2, 0
449	vperm	5, 6, 5, 31
450	vxor	3, 3, 1
451	bdnz	.Ldec_loop
452
453
454	addi	10, 11, 0x80
455
456	vperm	4, 14, 7, 2
457
458	lvx	2, 12, 10
459	vperm	1, 15, 7, 3
460	vxor	4, 4, 5
461	vxor	0, 1, 4
462	vperm	0, 0, 7, 2
463	blr
464.long	0
465.byte	0,12,0x14,0,0,0,0,0
466
467.globl	vpaes_decrypt
468.type	vpaes_decrypt,@function
469.align	5
470vpaes_decrypt:
471.localentry	vpaes_decrypt,0
472
473	stdu	1,-256(1)
474	li	10,63
475	li	11,79
476	mflr	6
477	li	7,-1
478	stvx	20,10,1
479	addi	10,10,32
480	stvx	21,11,1
481	addi	11,11,32
482	stvx	22,10,1
483	addi	10,10,32
484	stvx	23,11,1
485	addi	11,11,32
486	stvx	24,10,1
487	addi	10,10,32
488	stvx	25,11,1
489	addi	11,11,32
490	stvx	26,10,1
491	addi	10,10,32
492	stvx	27,11,1
493	addi	11,11,32
494	stvx	28,10,1
495	addi	10,10,32
496	stvx	29,11,1
497	addi	11,11,32
498	stvx	30,10,1
499	stvx	31,11,1
500	stw	7,252(1)
501	li	0, -1
502	std	6,272(1)
503	or	0,0,0
504
505	bl	_vpaes_decrypt_preheat
506
507	lvsr	27, 0, 3
508	lvx	0, 0, 3
509	addi	3, 3, 15
510	lvsl	29, 0, 4
511	lvsr	31, 0, 5
512	lvx	26, 0, 3
513	vperm	0, 26, 0, 27
514
515	bl	_vpaes_decrypt_core
516
517	andi.	8, 4, 15
518	li	9, 16
519	beq	.Ldec_out_aligned
520
521	vperm	0, 0, 0, 29
522	mtctr	9
523.Ldec_out_unaligned:
524	stvebx	0, 0, 4
525	addi	4, 4, 1
526	bdnz	.Ldec_out_unaligned
527	b	.Ldec_done
528
529.align	4
530.Ldec_out_aligned:
531	stvx	0, 0, 4
532.Ldec_done:
533
534	li	10,63
535	li	11,79
536	mtlr	6
537	or	7,7,7
538	lvx	20,10,1
539	addi	10,10,32
540	lvx	21,11,1
541	addi	11,11,32
542	lvx	22,10,1
543	addi	10,10,32
544	lvx	23,11,1
545	addi	11,11,32
546	lvx	24,10,1
547	addi	10,10,32
548	lvx	25,11,1
549	addi	11,11,32
550	lvx	26,10,1
551	addi	10,10,32
552	lvx	27,11,1
553	addi	11,11,32
554	lvx	28,10,1
555	addi	10,10,32
556	lvx	29,11,1
557	addi	11,11,32
558	lvx	30,10,1
559	lvx	31,11,1
560	addi	1,1,256
561	blr
562.long	0
563.byte	0,12,0x04,1,0x80,0,3,0
564.long	0
565.size	vpaes_decrypt,.-vpaes_decrypt
566
567.globl	vpaes_cbc_encrypt
568.type	vpaes_cbc_encrypt,@function
569.align	5
570vpaes_cbc_encrypt:
571.localentry	vpaes_cbc_encrypt,0
572
573	cmpldi	5,16
574	.long	0x4dc00020
575
576	stdu	1,-272(1)
577	mflr	0
578	li	10,63
579	li	11,79
580	li	12,-1
581	stvx	20,10,1
582	addi	10,10,32
583	stvx	21,11,1
584	addi	11,11,32
585	stvx	22,10,1
586	addi	10,10,32
587	stvx	23,11,1
588	addi	11,11,32
589	stvx	24,10,1
590	addi	10,10,32
591	stvx	25,11,1
592	addi	11,11,32
593	stvx	26,10,1
594	addi	10,10,32
595	stvx	27,11,1
596	addi	11,11,32
597	stvx	28,10,1
598	addi	10,10,32
599	stvx	29,11,1
600	addi	11,11,32
601	stvx	30,10,1
602	stvx	31,11,1
603	stw	12,252(1)
604	std	30,256(1)
605	std	31,264(1)
606	li	9, -16
607	std	0, 288(1)
608
609	and	30, 5, 9
610	andi.	9, 4, 15
611	mr	5, 6
612	mr	31, 7
613	li	6, -1
614	mcrf	1, 0
615	mr	7, 12
616	or	6,6,6
617
618	lvx	24, 0, 31
619	li	9, 15
620	lvsr	27, 0, 31
621	lvx	25, 9, 31
622	vperm	24, 25, 24, 27
623
624	cmpwi	8, 0
625	neg	8, 3
626	vxor	7, 7, 7
627	lvsr	31, 0, 5
628	lvsl	29, 0, 4
629	lvsl	27, 0, 8
630	vnor	30, 7, 7
631	lvx	26, 0, 3
632	vperm	30, 30, 7, 29
633	addi	3, 3, 15
634
635	beq	.Lcbc_decrypt
636
637	bl	_vpaes_encrypt_preheat
638	li	0, 16
639
640	beq	1, .Lcbc_enc_loop
641
642	vor	0,26,26
643	lvx	26, 0, 3
644	addi	3, 3, 16
645	vperm	0, 26, 0, 27
646	vxor	0, 0, 24
647
648	bl	_vpaes_encrypt_core
649
650	andi.	8, 4, 15
651	vor	24,0,0
652	sub	9, 4, 8
653	vperm	28, 0, 0, 29
654
655.Lcbc_enc_head:
656	stvebx	28, 8, 9
657	cmpwi	8, 15
658	addi	8, 8, 1
659	bne	.Lcbc_enc_head
660
661	sub.	30, 30, 0
662	addi	4, 4, 16
663	beq	.Lcbc_unaligned_done
664
665.Lcbc_enc_loop:
666	vor	0,26,26
667	lvx	26, 0, 3
668	addi	3, 3, 16
669	vperm	0, 26, 0, 27
670	vxor	0, 0, 24
671
672	bl	_vpaes_encrypt_core
673
674	vor	24,0,0
675	sub.	30, 30, 0
676	vperm	0, 0, 0, 29
677	vsel	1,28,0,30
678	vor	28,0,0
679	stvx	1, 0, 4
680	addi	4, 4, 16
681	bne	.Lcbc_enc_loop
682
683	b	.Lcbc_done
684
685.align	5
686.Lcbc_decrypt:
687	bl	_vpaes_decrypt_preheat
688	li	0, 16
689
690	beq	1, .Lcbc_dec_loop
691
692	vor	0,26,26
693	lvx	26, 0, 3
694	addi	3, 3, 16
695	vperm	0, 26, 0, 27
696	vor	25,0,0
697
698	bl	_vpaes_decrypt_core
699
700	andi.	8, 4, 15
701	vxor	0, 0, 24
702	vor	24,25,25
703	sub	9, 4, 8
704	vperm	28, 0, 0, 29
705
706.Lcbc_dec_head:
707	stvebx	28, 8, 9
708	cmpwi	8, 15
709	addi	8, 8, 1
710	bne	.Lcbc_dec_head
711
712	sub.	30, 30, 0
713	addi	4, 4, 16
714	beq	.Lcbc_unaligned_done
715
716.Lcbc_dec_loop:
717	vor	0,26,26
718	lvx	26, 0, 3
719	addi	3, 3, 16
720	vperm	0, 26, 0, 27
721	vor	25,0,0
722
723	bl	_vpaes_decrypt_core
724
725	vxor	0, 0, 24
726	vor	24,25,25
727	sub.	30, 30, 0
728	vperm	0, 0, 0, 29
729	vsel	1,28,0,30
730	vor	28,0,0
731	stvx	1, 0, 4
732	addi	4, 4, 16
733	bne	.Lcbc_dec_loop
734
735.Lcbc_done:
736	beq	1, .Lcbc_write_iv
737
738.Lcbc_unaligned_done:
739	andi.	8, 4, 15
740	sub	4, 4, 8
741	li	9, 0
742.Lcbc_tail:
743	stvebx	28, 9, 4
744	addi	9, 9, 1
745	cmpw	9, 8
746	bne	.Lcbc_tail
747
748.Lcbc_write_iv:
749	neg	8, 31
750	li	10, 4
751	lvsr	29, 0, 8
752	li	11, 8
753	li	12, 12
754	vperm	24, 24, 24, 29
755	stvewx	24, 0, 31
756	stvewx	24, 10, 31
757	stvewx	24, 11, 31
758	stvewx	24, 12, 31
759
760	or	7,7,7
761	li	10,63
762	li	11,79
763	lvx	20,10,1
764	addi	10,10,32
765	lvx	21,11,1
766	addi	11,11,32
767	lvx	22,10,1
768	addi	10,10,32
769	lvx	23,11,1
770	addi	11,11,32
771	lvx	24,10,1
772	addi	10,10,32
773	lvx	25,11,1
774	addi	11,11,32
775	lvx	26,10,1
776	addi	10,10,32
777	lvx	27,11,1
778	addi	11,11,32
779	lvx	28,10,1
780	addi	10,10,32
781	lvx	29,11,1
782	addi	11,11,32
783	lvx	30,10,1
784	lvx	31,11,1
785.Lcbc_abort:
786	ld	0, 288(1)
787	ld	30,256(1)
788	ld	31,264(1)
789	mtlr	0
790	addi	1,1,272
791	blr
792.long	0
793.byte	0,12,0x04,1,0x80,2,6,0
794.long	0
795.size	vpaes_cbc_encrypt,.-vpaes_cbc_encrypt
796
797
798
799
800
801.align	4
802_vpaes_key_preheat:
803	mflr	8
804	bl	.Lconsts
805	mtlr	8
806	li	11, 0xc0
807	li	10, 0xd0
808	li	9,  0xe0
809	li	8,  0xf0
810
811	vspltisb	8,4
812	vxor	9,9,9
813	lvx	10, 12, 11
814	li	11, 0x120
815	lvx	11, 12, 10
816	li	10, 0x130
817	lvx	12, 12, 9
818	li	9, 0x220
819	lvx	13, 12, 8
820	li	8, 0x230
821
822	lvx	14, 12, 11
823	li	11, 0x240
824	lvx	15, 12, 10
825	li	10, 0x250
826
827	lvx	16, 12, 9
828	li	9, 0x260
829	lvx	17, 12, 8
830	li	8, 0x270
831	lvx	18, 12, 11
832	li	11, 0x280
833	lvx	19, 12, 10
834	li	10, 0x290
835	lvx	20, 12, 9
836	li	9, 0x2a0
837	lvx	21, 12, 8
838	li	8, 0x2b0
839	lvx	22, 12, 11
840	lvx	23, 12, 10
841
842	lvx	24, 12, 9
843	lvx	25, 0, 12
844	lvx	26, 12, 8
845	blr
846.long	0
847.byte	0,12,0x14,0,0,0,0,0
848
849.align	4
850_vpaes_schedule_core:
851	mflr	7
852
853	bl	_vpaes_key_preheat
854
855
856	neg	8, 3
857	lvx	0, 0, 3
858	addi	3, 3, 15
859	lvsl	27, 0, 8
860	lvx	6, 0, 3
861	addi	3, 3, 8
862	vperm	0, 6, 0, 27
863
864
865	vor	3,0,0
866	bl	_vpaes_schedule_transform
867	vor	7,0,0
868
869	bne	1, .Lschedule_am_decrypting
870
871
872	li	8, 0x30
873	li	9, 4
874	li	10, 8
875	li	11, 12
876
877	lvsl	29, 0, 5
878	vnor	30, 9, 9
879	vperm	30, 30, 9, 29
880
881
882	vperm	28, 0, 0, 29
883	stvewx	28, 0, 5
884	stvewx	28, 9, 5
885	stvewx	28, 10, 5
886	addi	10, 12, 0x80
887	stvewx	28, 11, 5
888	b	.Lschedule_go
889
890.Lschedule_am_decrypting:
891	srwi	8, 4, 1
892	andi.	8, 8, 32
893	xori	8, 8, 32
894	addi	10, 12, 0x80
895
896	lvx	1, 8, 10
897	li	9, 4
898	li	10, 8
899	li	11, 12
900	vperm	4, 3, 3, 1
901
902	neg	0, 5
903	lvsr	29, 0, 0
904	vnor	30, 9, 9
905	vperm	30, 9, 30, 29
906
907
908	vperm	28, 4, 4, 29
909	stvewx	28, 0, 5
910	stvewx	28, 9, 5
911	stvewx	28, 10, 5
912	addi	10, 12, 0x80
913	stvewx	28, 11, 5
914	addi	5, 5, 15
915	xori	8, 8, 0x30
916
917.Lschedule_go:
918	cmplwi	4, 192
919	bgt	.Lschedule_256
920	beq	.Lschedule_192
921
922
923
924
925
926
927
928
929
930
931.Lschedule_128:
932	li	0, 10
933	mtctr	0
934
935.Loop_schedule_128:
936	bl	_vpaes_schedule_round
937	bdz	.Lschedule_mangle_last
938	bl	_vpaes_schedule_mangle
939	b	.Loop_schedule_128
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956.align	4
957.Lschedule_192:
958	li	0, 4
959	lvx	0, 0, 3
960	vperm	0, 0, 6, 27
961	vsldoi	0, 0, 3, 16-8
962	bl	_vpaes_schedule_transform
963	vsldoi	6, 9, 0, 16-8
964	vsldoi	6, 6, 9, 16-8
965	mtctr	0
966
967.Loop_schedule_192:
968	bl	_vpaes_schedule_round
969	vsldoi	0, 0, 6, 16-8
970	bl	_vpaes_schedule_mangle
971	bl	_vpaes_schedule_192_smear
972	bl	_vpaes_schedule_mangle
973	bl	_vpaes_schedule_round
974	bdz	.Lschedule_mangle_last
975	bl	_vpaes_schedule_mangle
976	bl	_vpaes_schedule_192_smear
977	b	.Loop_schedule_192
978
979
980
981
982
983
984
985
986
987
988
989.align	4
990.Lschedule_256:
991	li	0, 7
992	addi	3, 3, 8
993	lvx	0, 0, 3
994	vperm	0, 0, 6, 27
995	bl	_vpaes_schedule_transform
996	mtctr	0
997
998.Loop_schedule_256:
999	bl	_vpaes_schedule_mangle
1000	vor	6,0,0
1001
1002
1003	bl	_vpaes_schedule_round
1004	bdz	.Lschedule_mangle_last
1005	bl	_vpaes_schedule_mangle
1006
1007
1008	vspltw	0, 0, 3-3
1009	vor	5,7,7
1010	vor	7,6,6
1011	bl	_vpaes_schedule_low_round
1012	vor	7,5,5
1013
1014	b	.Loop_schedule_256
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025.align	4
1026.Lschedule_mangle_last:
1027
1028	li	11, 0x2e0
1029	li	9,  0x2f0
1030	bne	1, .Lschedule_mangle_last_dec
1031
1032
1033	lvx	1, 8, 10
1034	li	11, 0x2c0
1035	li	9,  0x2d0
1036	vperm	0, 0, 0, 1
1037
1038	lvx	12, 11, 12
1039	lvx	13, 9, 12
1040	addi	5, 5, 16
1041	vxor	0, 0, 26
1042	bl	_vpaes_schedule_transform
1043
1044
1045	vperm	0, 0, 0, 29
1046	li	10, 4
1047	vsel	2,28,0,30
1048	li	11, 8
1049	stvx	2, 0, 5
1050	li	12, 12
1051	stvewx	0, 0, 5
1052	stvewx	0, 10, 5
1053	stvewx	0, 11, 5
1054	stvewx	0, 12, 5
1055	b	.Lschedule_mangle_done
1056
1057.align	4
1058.Lschedule_mangle_last_dec:
1059	lvx	12, 11, 12
1060	lvx	13, 9,  12
1061	addi	5, 5, -16
1062	vxor	0, 0, 26
1063	bl	_vpaes_schedule_transform
1064
1065
1066	addi	9, 5, -15
1067	vperm	0, 0, 0, 29
1068	li	10, 4
1069	vsel	2,28,0,30
1070	li	11, 8
1071	stvx	2, 0, 5
1072	li	12, 12
1073	stvewx	0, 0, 9
1074	stvewx	0, 10, 9
1075	stvewx	0, 11, 9
1076	stvewx	0, 12, 9
1077
1078
1079.Lschedule_mangle_done:
1080	mtlr	7
1081
1082	vxor	0, 0, 0
1083	vxor	1, 1, 1
1084	vxor	2, 2, 2
1085	vxor	3, 3, 3
1086	vxor	4, 4, 4
1087	vxor	5, 5, 5
1088	vxor	6, 6, 6
1089	vxor	7, 7, 7
1090
1091	blr
1092.long	0
1093.byte	0,12,0x14,0,0,0,0,0
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109.align	4
1110_vpaes_schedule_192_smear:
1111	vspltw	0, 7, 3-3
1112	vsldoi	1, 6, 9, 16-12
1113	vsldoi	0, 0, 7, 16-8
1114	vxor	6, 6, 1
1115	vxor	6, 6, 0
1116	vor	0,6,6
1117	vsldoi	6, 9, 6, 16-8
1118	vsldoi	6, 6, 9, 16-8
1119	blr
1120.long	0
1121.byte	0,12,0x14,0,0,0,0,0
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141.align	4
1142_vpaes_schedule_round:
1143
1144
1145	vsldoi	1, 9, 24, 16-15
1146	vsldoi	24, 24, 24, 16-15
1147	vxor	7, 7, 1
1148
1149
1150	vspltw	0, 0, 3-3
1151	vsldoi	0, 0, 0, 16-1
1152
1153
1154
1155
1156_vpaes_schedule_low_round:
1157
1158	vsldoi	1, 7, 9, 16-12
1159	vxor	7, 7, 1
1160	vspltisb	1,0x0f
1161	vsldoi	4, 7, 9, 16-8
1162
1163
1164	vand	1, 1, 0
1165	vsrb	0, 0, 8
1166	vxor	7, 7, 4
1167	vperm	2, 11, 9, 1
1168	vxor	1, 1, 0
1169	vperm	3, 10, 9, 0
1170	vxor	3, 3, 2
1171	vperm	4, 10, 9, 1
1172	vxor	7, 7, 26
1173	vperm	3, 10, 9, 3
1174	vxor	4, 4, 2
1175	vperm	2, 10, 9, 4
1176	vxor	3, 3, 1
1177	vxor	2, 2, 0
1178	vperm	4, 15, 9, 3
1179	vperm	1, 14, 9, 2
1180	vxor	1, 1, 4
1181
1182
1183	vxor	0, 1, 7
1184	vxor	7, 1, 7
1185	blr
1186.long	0
1187.byte	0,12,0x14,0,0,0,0,0
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198.align	4
1199_vpaes_schedule_transform:
1200
1201	vsrb	2, 0, 8
1202
1203	vperm	0, 12, 12, 0
1204
1205	vperm	2, 13, 13, 2
1206	vxor	0, 0, 2
1207	blr
1208.long	0
1209.byte	0,12,0x14,0,0,0,0,0
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234.align	4
1235_vpaes_schedule_mangle:
1236
1237
1238	bne	1, .Lschedule_mangle_dec
1239
1240
1241	vxor	4, 0, 26
1242	addi	5, 5, 16
1243	vperm	4, 4, 4, 25
1244	vperm	1, 4, 4, 25
1245	vperm	3, 1, 1, 25
1246	vxor	4, 4, 1
1247	lvx	1, 8, 10
1248	vxor	3, 3, 4
1249
1250	vperm	3, 3, 3, 1
1251	addi	8, 8, -16
1252	andi.	8, 8, 0x30
1253
1254
1255	vperm	1, 3, 3, 29
1256	vsel	2,28,1,30
1257	vor	28,1,1
1258	stvx	2, 0, 5
1259	blr
1260
1261.align	4
1262.Lschedule_mangle_dec:
1263
1264
1265	vsrb	1, 0, 8
1266
1267
1268
1269	vperm	2, 16, 16, 0
1270
1271	vperm	3, 17, 17, 1
1272	vxor	3, 3, 2
1273	vperm	3, 3, 9, 25
1274
1275
1276	vperm	2, 18, 18, 0
1277	vxor	2, 2, 3
1278
1279	vperm	3, 19, 19, 1
1280	vxor	3, 3, 2
1281	vperm	3, 3, 9, 25
1282
1283
1284	vperm	2, 20, 20, 0
1285	vxor	2, 2, 3
1286
1287	vperm	3, 21, 21, 1
1288	vxor	3, 3, 2
1289
1290
1291	vperm	2, 22, 22, 0
1292	vperm	3, 3, 9, 25
1293
1294	vperm	4, 23, 23, 1
1295	lvx	1, 8, 10
1296	vxor	2, 2, 3
1297	vxor	3, 4, 2
1298
1299	addi	5, 5, -16
1300
1301	vperm	3, 3, 3, 1
1302	addi	8, 8, -16
1303	andi.	8, 8, 0x30
1304
1305
1306	vperm	1, 3, 3, 29
1307	vsel	2,28,1,30
1308	vor	28,1,1
1309	stvx	2, 0, 5
1310	blr
1311.long	0
1312.byte	0,12,0x14,0,0,0,0,0
1313
1314.globl	vpaes_set_encrypt_key
1315.type	vpaes_set_encrypt_key,@function
1316.align	5
1317vpaes_set_encrypt_key:
1318.localentry	vpaes_set_encrypt_key,0
1319
1320	stdu	1,-256(1)
1321	li	10,63
1322	li	11,79
1323	mflr	0
1324	li	6,-1
1325	stvx	20,10,1
1326	addi	10,10,32
1327	stvx	21,11,1
1328	addi	11,11,32
1329	stvx	22,10,1
1330	addi	10,10,32
1331	stvx	23,11,1
1332	addi	11,11,32
1333	stvx	24,10,1
1334	addi	10,10,32
1335	stvx	25,11,1
1336	addi	11,11,32
1337	stvx	26,10,1
1338	addi	10,10,32
1339	stvx	27,11,1
1340	addi	11,11,32
1341	stvx	28,10,1
1342	addi	10,10,32
1343	stvx	29,11,1
1344	addi	11,11,32
1345	stvx	30,10,1
1346	stvx	31,11,1
1347	stw	6,252(1)
1348	li	7, -1
1349	std	0, 272(1)
1350	or	7,7,7
1351
1352	srwi	9, 4, 5
1353	addi	9, 9, 6
1354	stw	9, 240(5)
1355
1356	cmplw	1,4,4
1357	li	8, 0x30
1358	bl	_vpaes_schedule_core
1359
1360	ld	0, 272(1)
1361	li	10,63
1362	li	11,79
1363	or	6,6,6
1364	mtlr	0
1365	xor	3, 3, 3
1366	lvx	20,10,1
1367	addi	10,10,32
1368	lvx	21,11,1
1369	addi	11,11,32
1370	lvx	22,10,1
1371	addi	10,10,32
1372	lvx	23,11,1
1373	addi	11,11,32
1374	lvx	24,10,1
1375	addi	10,10,32
1376	lvx	25,11,1
1377	addi	11,11,32
1378	lvx	26,10,1
1379	addi	10,10,32
1380	lvx	27,11,1
1381	addi	11,11,32
1382	lvx	28,10,1
1383	addi	10,10,32
1384	lvx	29,11,1
1385	addi	11,11,32
1386	lvx	30,10,1
1387	lvx	31,11,1
1388	addi	1,1,256
1389	blr
1390.long	0
1391.byte	0,12,0x04,1,0x80,0,3,0
1392.long	0
1393.size	vpaes_set_encrypt_key,.-vpaes_set_encrypt_key
1394
1395.globl	vpaes_set_decrypt_key
1396.type	vpaes_set_decrypt_key,@function
1397.align	4
1398vpaes_set_decrypt_key:
1399.localentry	vpaes_set_decrypt_key,0
1400
1401	stdu	1,-256(1)
1402	li	10,63
1403	li	11,79
1404	mflr	0
1405	li	6,-1
1406	stvx	20,10,1
1407	addi	10,10,32
1408	stvx	21,11,1
1409	addi	11,11,32
1410	stvx	22,10,1
1411	addi	10,10,32
1412	stvx	23,11,1
1413	addi	11,11,32
1414	stvx	24,10,1
1415	addi	10,10,32
1416	stvx	25,11,1
1417	addi	11,11,32
1418	stvx	26,10,1
1419	addi	10,10,32
1420	stvx	27,11,1
1421	addi	11,11,32
1422	stvx	28,10,1
1423	addi	10,10,32
1424	stvx	29,11,1
1425	addi	11,11,32
1426	stvx	30,10,1
1427	stvx	31,11,1
1428	stw	6,252(1)
1429	li	7, -1
1430	std	0, 272(1)
1431	or	7,7,7
1432
1433	srwi	9, 4, 5
1434	addi	9, 9, 6
1435	stw	9, 240(5)
1436
1437	slwi	9, 9, 4
1438	add	5, 5, 9
1439
1440	cmplwi	1, 4, 0
1441	srwi	8, 4, 1
1442	andi.	8, 8, 32
1443	xori	8, 8, 32
1444	bl	_vpaes_schedule_core
1445
1446	ld	0,  272(1)
1447	li	10,63
1448	li	11,79
1449	or	6,6,6
1450	mtlr	0
1451	xor	3, 3, 3
1452	lvx	20,10,1
1453	addi	10,10,32
1454	lvx	21,11,1
1455	addi	11,11,32
1456	lvx	22,10,1
1457	addi	10,10,32
1458	lvx	23,11,1
1459	addi	11,11,32
1460	lvx	24,10,1
1461	addi	10,10,32
1462	lvx	25,11,1
1463	addi	11,11,32
1464	lvx	26,10,1
1465	addi	10,10,32
1466	lvx	27,11,1
1467	addi	11,11,32
1468	lvx	28,10,1
1469	addi	10,10,32
1470	lvx	29,11,1
1471	addi	11,11,32
1472	lvx	30,10,1
1473	lvx	31,11,1
1474	addi	1,1,256
1475	blr
1476.long	0
1477.byte	0,12,0x04,1,0x80,0,3,0
1478.long	0
1479.size	vpaes_set_decrypt_key,.-vpaes_set_decrypt_key
1480