CoCalc -- keccak1600-armv8.S

GitHub Repository: freebsd/freebsd-src
Path: blob/main/sys/crypto/openssl/aarch64/keccak1600-armv8.S
³⁹⁵⁰⁷ views
1
/* Do not modify. This file is auto-generated from keccak1600-armv8.pl. */
2
#include "arm_arch.h"
3

4
.section	.rodata
5

6
.align	8	// strategic alignment and padding that allows to use
7
		// address value as loop termination condition...
8
.quad	0,0,0,0,0,0,0,0
9
.type	iotas,%object
10
iotas:
11
.quad	0x0000000000000001
12
.quad	0x0000000000008082
13
.quad	0x800000000000808a
14
.quad	0x8000000080008000
15
.quad	0x000000000000808b
16
.quad	0x0000000080000001
17
.quad	0x8000000080008081
18
.quad	0x8000000000008009
19
.quad	0x000000000000008a
20
.quad	0x0000000000000088
21
.quad	0x0000000080008009
22
.quad	0x000000008000000a
23
.quad	0x000000008000808b
24
.quad	0x800000000000008b
25
.quad	0x8000000000008089
26
.quad	0x8000000000008003
27
.quad	0x8000000000008002
28
.quad	0x8000000000000080
29
.quad	0x000000000000800a
30
.quad	0x800000008000000a
31
.quad	0x8000000080008081
32
.quad	0x8000000000008080
33
.quad	0x0000000080000001
34
.quad	0x8000000080008008
35
.size	iotas,.-iotas
36
.text
37

38
.type	KeccakF1600_int,%function
39
.align	5
40
KeccakF1600_int:
41
	AARCH64_SIGN_LINK_REGISTER
42
	adrp	x28,iotas
43
	add	x28,x28,#:lo12:iotas
44
	stp	x28,x30,[sp,#16]		// 32 bytes on top are mine
45
	b	.Loop
46
.align	4
47
.Loop:
48
	////////////////////////////////////////// Theta
49
	eor	x26,x0,x5
50
	stp	x4,x9,[sp,#0]	// offload pair...
51
	eor	x27,x1,x6
52
	eor	x28,x2,x7
53
	eor	x30,x3,x8
54
	eor	x4,x4,x9
55
	eor	x26,x26,x10
56
	eor	x27,x27,x11
57
	eor	x28,x28,x12
58
	eor	x30,x30,x13
59
	eor	x4,x4,x14
60
	eor	x26,x26,x15
61
	eor	x27,x27,x16
62
	eor	x28,x28,x17
63
	eor	x30,x30,x25
64
	eor	x4,x4,x19
65
	eor	x26,x26,x20
66
	eor	x28,x28,x22
67
	eor	x27,x27,x21
68
	eor	x30,x30,x23
69
	eor	x4,x4,x24
70

71
	eor	x9,x26,x28,ror#63
72

73
	eor	x1,x1,x9
74
	eor	x6,x6,x9
75
	eor	x11,x11,x9
76
	eor	x16,x16,x9
77
	eor	x21,x21,x9
78

79
	eor	x9,x27,x30,ror#63
80
	eor	x28,x28,x4,ror#63
81
	eor	x30,x30,x26,ror#63
82
	eor	x4,x4,x27,ror#63
83

84
	eor	x27,   x2,x9		// mov	x27,x2
85
	eor	x7,x7,x9
86
	eor	x12,x12,x9
87
	eor	x17,x17,x9
88
	eor	x22,x22,x9
89

90
	eor	x0,x0,x4
91
	eor	x5,x5,x4
92
	eor	x10,x10,x4
93
	eor	x15,x15,x4
94
	eor	x20,x20,x4
95
	ldp	x4,x9,[sp,#0]	// re-load offloaded data
96
	eor	x26,   x3,x28		// mov	x26,x3
97
	eor	x8,x8,x28
98
	eor	x13,x13,x28
99
	eor	x25,x25,x28
100
	eor	x23,x23,x28
101

102
	eor	x28,   x4,x30		// mov	x28,x4
103
	eor	x9,x9,x30
104
	eor	x14,x14,x30
105
	eor	x19,x19,x30
106
	eor	x24,x24,x30
107

108
	////////////////////////////////////////// Rho+Pi
109
	mov	x30,x1
110
	ror	x1,x6,#64-44
111
	//mov	x27,x2
112
	ror	x2,x12,#64-43
113
	//mov	x26,x3
114
	ror	x3,x25,#64-21
115
	//mov	x28,x4
116
	ror	x4,x24,#64-14
117

118
	ror	x6,x9,#64-20
119
	ror	x12,x13,#64-25
120
	ror	x25,x17,#64-15
121
	ror	x24,x21,#64-2
122

123
	ror	x9,x22,#64-61
124
	ror	x13,x19,#64-8
125
	ror	x17,x11,#64-10
126
	ror	x21,x8,#64-55
127

128
	ror	x22,x14,#64-39
129
	ror	x19,x23,#64-56
130
	ror	x11,x7,#64-6
131
	ror	x8,x16,#64-45
132

133
	ror	x14,x20,#64-18
134
	ror	x23,x15,#64-41
135
	ror	x7,x10,#64-3
136
	ror	x16,x5,#64-36
137

138
	ror	x5,x26,#64-28
139
	ror	x10,x30,#64-1
140
	ror	x15,x28,#64-27
141
	ror	x20,x27,#64-62
142

143
	////////////////////////////////////////// Chi+Iota
144
	bic	x26,x2,x1
145
	bic	x27,x3,x2
146
	bic	x28,x0,x4
147
	bic	x30,x1,x0
148
	eor	x0,x0,x26
149
	bic	x26,x4,x3
150
	eor	x1,x1,x27
151
	ldr	x27,[sp,#16]
152
	eor	x3,x3,x28
153
	eor	x4,x4,x30
154
	eor	x2,x2,x26
155
	ldr	x30,[x27],#8		// Iota[i++]
156

157
	bic	x26,x7,x6
158
	tst	x27,#255			// are we done?
159
	str	x27,[sp,#16]
160
	bic	x27,x8,x7
161
	bic	x28,x5,x9
162
	eor	x0,x0,x30		// A[0][0] ^= Iota
163
	bic	x30,x6,x5
164
	eor	x5,x5,x26
165
	bic	x26,x9,x8
166
	eor	x6,x6,x27
167
	eor	x8,x8,x28
168
	eor	x9,x9,x30
169
	eor	x7,x7,x26
170

171
	bic	x26,x12,x11
172
	bic	x27,x13,x12
173
	bic	x28,x10,x14
174
	bic	x30,x11,x10
175
	eor	x10,x10,x26
176
	bic	x26,x14,x13
177
	eor	x11,x11,x27
178
	eor	x13,x13,x28
179
	eor	x14,x14,x30
180
	eor	x12,x12,x26
181

182
	bic	x26,x17,x16
183
	bic	x27,x25,x17
184
	bic	x28,x15,x19
185
	bic	x30,x16,x15
186
	eor	x15,x15,x26
187
	bic	x26,x19,x25
188
	eor	x16,x16,x27
189
	eor	x25,x25,x28
190
	eor	x19,x19,x30
191
	eor	x17,x17,x26
192

193
	bic	x26,x22,x21
194
	bic	x27,x23,x22
195
	bic	x28,x20,x24
196
	bic	x30,x21,x20
197
	eor	x20,x20,x26
198
	bic	x26,x24,x23
199
	eor	x21,x21,x27
200
	eor	x23,x23,x28
201
	eor	x24,x24,x30
202
	eor	x22,x22,x26
203

204
	bne	.Loop
205

206
	ldr	x30,[sp,#24]
207
	AARCH64_VALIDATE_LINK_REGISTER
208
	ret
209
.size	KeccakF1600_int,.-KeccakF1600_int
210

211
.type	KeccakF1600,%function
212
.align	5
213
KeccakF1600:
214
	AARCH64_SIGN_LINK_REGISTER
215
	stp	x29,x30,[sp,#-128]!
216
	add	x29,sp,#0
217
	stp	x19,x20,[sp,#16]
218
	stp	x21,x22,[sp,#32]
219
	stp	x23,x24,[sp,#48]
220
	stp	x25,x26,[sp,#64]
221
	stp	x27,x28,[sp,#80]
222
	sub	sp,sp,#48
223

224
	str	x0,[sp,#32]			// offload argument
225
	mov	x26,x0
226
	ldp	x0,x1,[x0,#16*0]
227
	ldp	x2,x3,[x26,#16*1]
228
	ldp	x4,x5,[x26,#16*2]
229
	ldp	x6,x7,[x26,#16*3]
230
	ldp	x8,x9,[x26,#16*4]
231
	ldp	x10,x11,[x26,#16*5]
232
	ldp	x12,x13,[x26,#16*6]
233
	ldp	x14,x15,[x26,#16*7]
234
	ldp	x16,x17,[x26,#16*8]
235
	ldp	x25,x19,[x26,#16*9]
236
	ldp	x20,x21,[x26,#16*10]
237
	ldp	x22,x23,[x26,#16*11]
238
	ldr	x24,[x26,#16*12]
239

240
	bl	KeccakF1600_int
241

242
	ldr	x26,[sp,#32]
243
	stp	x0,x1,[x26,#16*0]
244
	stp	x2,x3,[x26,#16*1]
245
	stp	x4,x5,[x26,#16*2]
246
	stp	x6,x7,[x26,#16*3]
247
	stp	x8,x9,[x26,#16*4]
248
	stp	x10,x11,[x26,#16*5]
249
	stp	x12,x13,[x26,#16*6]
250
	stp	x14,x15,[x26,#16*7]
251
	stp	x16,x17,[x26,#16*8]
252
	stp	x25,x19,[x26,#16*9]
253
	stp	x20,x21,[x26,#16*10]
254
	stp	x22,x23,[x26,#16*11]
255
	str	x24,[x26,#16*12]
256

257
	ldp	x19,x20,[x29,#16]
258
	add	sp,sp,#48
259
	ldp	x21,x22,[x29,#32]
260
	ldp	x23,x24,[x29,#48]
261
	ldp	x25,x26,[x29,#64]
262
	ldp	x27,x28,[x29,#80]
263
	ldp	x29,x30,[sp],#128
264
	AARCH64_VALIDATE_LINK_REGISTER
265
	ret
266
.size	KeccakF1600,.-KeccakF1600
267

268
.globl	SHA3_absorb
269
.type	SHA3_absorb,%function
270
.align	5
271
SHA3_absorb:
272
	AARCH64_SIGN_LINK_REGISTER
273
	stp	x29,x30,[sp,#-128]!
274
	add	x29,sp,#0
275
	stp	x19,x20,[sp,#16]
276
	stp	x21,x22,[sp,#32]
277
	stp	x23,x24,[sp,#48]
278
	stp	x25,x26,[sp,#64]
279
	stp	x27,x28,[sp,#80]
280
	sub	sp,sp,#64
281

282
	stp	x0,x1,[sp,#32]			// offload arguments
283
	stp	x2,x3,[sp,#48]
284

285
	mov	x26,x0			// uint64_t A[5][5]
286
	mov	x27,x1			// const void *inp
287
	mov	x28,x2			// size_t len
288
	mov	x30,x3			// size_t bsz
289
	ldp	x0,x1,[x26,#16*0]
290
	ldp	x2,x3,[x26,#16*1]
291
	ldp	x4,x5,[x26,#16*2]
292
	ldp	x6,x7,[x26,#16*3]
293
	ldp	x8,x9,[x26,#16*4]
294
	ldp	x10,x11,[x26,#16*5]
295
	ldp	x12,x13,[x26,#16*6]
296
	ldp	x14,x15,[x26,#16*7]
297
	ldp	x16,x17,[x26,#16*8]
298
	ldp	x25,x19,[x26,#16*9]
299
	ldp	x20,x21,[x26,#16*10]
300
	ldp	x22,x23,[x26,#16*11]
301
	ldr	x24,[x26,#16*12]
302
	b	.Loop_absorb
303

304
.align	4
305
.Loop_absorb:
306
	subs	x26,x28,x30		// len - bsz
307
	blo	.Labsorbed
308

309
	str	x26,[sp,#48]			// save len - bsz
310
	ldr	x26,[x27],#8		// *inp++
311
#ifdef	__AARCH64EB__
312
	rev	x26,x26
313
#endif
314
	eor	x0,x0,x26
315
	cmp	x30,#8*(0+2)
316
	blo	.Lprocess_block
317
	ldr	x26,[x27],#8		// *inp++
318
#ifdef	__AARCH64EB__
319
	rev	x26,x26
320
#endif
321
	eor	x1,x1,x26
322
	beq	.Lprocess_block
323
	ldr	x26,[x27],#8		// *inp++
324
#ifdef	__AARCH64EB__
325
	rev	x26,x26
326
#endif
327
	eor	x2,x2,x26
328
	cmp	x30,#8*(2+2)
329
	blo	.Lprocess_block
330
	ldr	x26,[x27],#8		// *inp++
331
#ifdef	__AARCH64EB__
332
	rev	x26,x26
333
#endif
334
	eor	x3,x3,x26
335
	beq	.Lprocess_block
336
	ldr	x26,[x27],#8		// *inp++
337
#ifdef	__AARCH64EB__
338
	rev	x26,x26
339
#endif
340
	eor	x4,x4,x26
341
	cmp	x30,#8*(4+2)
342
	blo	.Lprocess_block
343
	ldr	x26,[x27],#8		// *inp++
344
#ifdef	__AARCH64EB__
345
	rev	x26,x26
346
#endif
347
	eor	x5,x5,x26
348
	beq	.Lprocess_block
349
	ldr	x26,[x27],#8		// *inp++
350
#ifdef	__AARCH64EB__
351
	rev	x26,x26
352
#endif
353
	eor	x6,x6,x26
354
	cmp	x30,#8*(6+2)
355
	blo	.Lprocess_block
356
	ldr	x26,[x27],#8		// *inp++
357
#ifdef	__AARCH64EB__
358
	rev	x26,x26
359
#endif
360
	eor	x7,x7,x26
361
	beq	.Lprocess_block
362
	ldr	x26,[x27],#8		// *inp++
363
#ifdef	__AARCH64EB__
364
	rev	x26,x26
365
#endif
366
	eor	x8,x8,x26
367
	cmp	x30,#8*(8+2)
368
	blo	.Lprocess_block
369
	ldr	x26,[x27],#8		// *inp++
370
#ifdef	__AARCH64EB__
371
	rev	x26,x26
372
#endif
373
	eor	x9,x9,x26
374
	beq	.Lprocess_block
375
	ldr	x26,[x27],#8		// *inp++
376
#ifdef	__AARCH64EB__
377
	rev	x26,x26
378
#endif
379
	eor	x10,x10,x26
380
	cmp	x30,#8*(10+2)
381
	blo	.Lprocess_block
382
	ldr	x26,[x27],#8		// *inp++
383
#ifdef	__AARCH64EB__
384
	rev	x26,x26
385
#endif
386
	eor	x11,x11,x26
387
	beq	.Lprocess_block
388
	ldr	x26,[x27],#8		// *inp++
389
#ifdef	__AARCH64EB__
390
	rev	x26,x26
391
#endif
392
	eor	x12,x12,x26
393
	cmp	x30,#8*(12+2)
394
	blo	.Lprocess_block
395
	ldr	x26,[x27],#8		// *inp++
396
#ifdef	__AARCH64EB__
397
	rev	x26,x26
398
#endif
399
	eor	x13,x13,x26
400
	beq	.Lprocess_block
401
	ldr	x26,[x27],#8		// *inp++
402
#ifdef	__AARCH64EB__
403
	rev	x26,x26
404
#endif
405
	eor	x14,x14,x26
406
	cmp	x30,#8*(14+2)
407
	blo	.Lprocess_block
408
	ldr	x26,[x27],#8		// *inp++
409
#ifdef	__AARCH64EB__
410
	rev	x26,x26
411
#endif
412
	eor	x15,x15,x26
413
	beq	.Lprocess_block
414
	ldr	x26,[x27],#8		// *inp++
415
#ifdef	__AARCH64EB__
416
	rev	x26,x26
417
#endif
418
	eor	x16,x16,x26
419
	cmp	x30,#8*(16+2)
420
	blo	.Lprocess_block
421
	ldr	x26,[x27],#8		// *inp++
422
#ifdef	__AARCH64EB__
423
	rev	x26,x26
424
#endif
425
	eor	x17,x17,x26
426
	beq	.Lprocess_block
427
	ldr	x26,[x27],#8		// *inp++
428
#ifdef	__AARCH64EB__
429
	rev	x26,x26
430
#endif
431
	eor	x25,x25,x26
432
	cmp	x30,#8*(18+2)
433
	blo	.Lprocess_block
434
	ldr	x26,[x27],#8		// *inp++
435
#ifdef	__AARCH64EB__
436
	rev	x26,x26
437
#endif
438
	eor	x19,x19,x26
439
	beq	.Lprocess_block
440
	ldr	x26,[x27],#8		// *inp++
441
#ifdef	__AARCH64EB__
442
	rev	x26,x26
443
#endif
444
	eor	x20,x20,x26
445
	cmp	x30,#8*(20+2)
446
	blo	.Lprocess_block
447
	ldr	x26,[x27],#8		// *inp++
448
#ifdef	__AARCH64EB__
449
	rev	x26,x26
450
#endif
451
	eor	x21,x21,x26
452
	beq	.Lprocess_block
453
	ldr	x26,[x27],#8		// *inp++
454
#ifdef	__AARCH64EB__
455
	rev	x26,x26
456
#endif
457
	eor	x22,x22,x26
458
	cmp	x30,#8*(22+2)
459
	blo	.Lprocess_block
460
	ldr	x26,[x27],#8		// *inp++
461
#ifdef	__AARCH64EB__
462
	rev	x26,x26
463
#endif
464
	eor	x23,x23,x26
465
	beq	.Lprocess_block
466
	ldr	x26,[x27],#8		// *inp++
467
#ifdef	__AARCH64EB__
468
	rev	x26,x26
469
#endif
470
	eor	x24,x24,x26
471

472
.Lprocess_block:
473
	str	x27,[sp,#40]			// save inp
474

475
	bl	KeccakF1600_int
476

477
	ldr	x27,[sp,#40]			// restore arguments
478
	ldp	x28,x30,[sp,#48]
479
	b	.Loop_absorb
480

481
.align	4
482
.Labsorbed:
483
	ldr	x27,[sp,#32]
484
	stp	x0,x1,[x27,#16*0]
485
	stp	x2,x3,[x27,#16*1]
486
	stp	x4,x5,[x27,#16*2]
487
	stp	x6,x7,[x27,#16*3]
488
	stp	x8,x9,[x27,#16*4]
489
	stp	x10,x11,[x27,#16*5]
490
	stp	x12,x13,[x27,#16*6]
491
	stp	x14,x15,[x27,#16*7]
492
	stp	x16,x17,[x27,#16*8]
493
	stp	x25,x19,[x27,#16*9]
494
	stp	x20,x21,[x27,#16*10]
495
	stp	x22,x23,[x27,#16*11]
496
	str	x24,[x27,#16*12]
497

498
	mov	x0,x28			// return value
499
	ldp	x19,x20,[x29,#16]
500
	add	sp,sp,#64
501
	ldp	x21,x22,[x29,#32]
502
	ldp	x23,x24,[x29,#48]
503
	ldp	x25,x26,[x29,#64]
504
	ldp	x27,x28,[x29,#80]
505
	ldp	x29,x30,[sp],#128
506
	AARCH64_VALIDATE_LINK_REGISTER
507
	ret
508
.size	SHA3_absorb,.-SHA3_absorb
509
.globl	SHA3_squeeze
510
.type	SHA3_squeeze,%function
511
.align	5
512
SHA3_squeeze:
513
	AARCH64_SIGN_LINK_REGISTER
514
	stp	x29,x30,[sp,#-48]!
515
	add	x29,sp,#0
516
	stp	x19,x20,[sp,#16]
517
	stp	x21,x22,[sp,#32]
518

519
	mov	x19,x0			// put aside arguments
520
	mov	x20,x1
521
	mov	x21,x2
522
	mov	x22,x3
523
	cmp	w4, #0				// w4 = 'next' argument
524
	bne	.Lnext_block
525

526
.Loop_squeeze:
527
	ldr	x4,[x0],#8
528
	cmp	x21,#8
529
	blo	.Lsqueeze_tail
530
#ifdef	__AARCH64EB__
531
	rev	x4,x4
532
#endif
533
	str	x4,[x20],#8
534
	subs	x21,x21,#8
535
	beq	.Lsqueeze_done
536

537
	subs	x3,x3,#8
538
	bhi	.Loop_squeeze
539
.Lnext_block:
540
	mov	x0,x19
541
	bl	KeccakF1600
542
	mov	x0,x19
543
	mov	x3,x22
544
	b	.Loop_squeeze
545

546
.align	4
547
.Lsqueeze_tail:
548
	strb	w4,[x20],#1
549
	lsr	x4,x4,#8
550
	subs	x21,x21,#1
551
	beq	.Lsqueeze_done
552
	strb	w4,[x20],#1
553
	lsr	x4,x4,#8
554
	subs	x21,x21,#1
555
	beq	.Lsqueeze_done
556
	strb	w4,[x20],#1
557
	lsr	x4,x4,#8
558
	subs	x21,x21,#1
559
	beq	.Lsqueeze_done
560
	strb	w4,[x20],#1
561
	lsr	x4,x4,#8
562
	subs	x21,x21,#1
563
	beq	.Lsqueeze_done
564
	strb	w4,[x20],#1
565
	lsr	x4,x4,#8
566
	subs	x21,x21,#1
567
	beq	.Lsqueeze_done
568
	strb	w4,[x20],#1
569
	lsr	x4,x4,#8
570
	subs	x21,x21,#1
571
	beq	.Lsqueeze_done
572
	strb	w4,[x20],#1
573

574
.Lsqueeze_done:
575
	ldp	x19,x20,[sp,#16]
576
	ldp	x21,x22,[sp,#32]
577
	ldp	x29,x30,[sp],#48
578
	AARCH64_VALIDATE_LINK_REGISTER
579
	ret
580
.size	SHA3_squeeze,.-SHA3_squeeze
581
.type	KeccakF1600_ce,%function
582
.align	5
583
KeccakF1600_ce:
584
	mov	x9,#24
585
	adrp	x10,iotas
586
	add	x10,x10,#:lo12:iotas
587
	b	.Loop_ce
588
.align	4
589
.Loop_ce:
590
	////////////////////////////////////////////////// Theta
591
.inst	0xce0f2a99	//eor3 v25.16b,v20.16b,v15.16b,v10.16b
592
.inst	0xce102eba	//eor3 v26.16b,v21.16b,v16.16b,v11.16b
593
.inst	0xce1132db	//eor3 v27.16b,v22.16b,v17.16b,v12.16b
594
.inst	0xce1236fc	//eor3 v28.16b,v23.16b,v18.16b,v13.16b
595
.inst	0xce133b1d	//eor3 v29.16b,v24.16b,v19.16b,v14.16b
596
.inst	0xce050339	//eor3 v25.16b,v25.16b,   v5.16b,v0.16b
597
.inst	0xce06075a	//eor3 v26.16b,v26.16b,   v6.16b,v1.16b
598
.inst	0xce070b7b	//eor3 v27.16b,v27.16b,   v7.16b,v2.16b
599
.inst	0xce080f9c	//eor3 v28.16b,v28.16b,   v8.16b,v3.16b
600
.inst	0xce0913bd	//eor3 v29.16b,v29.16b,   v9.16b,v4.16b
601

602
.inst	0xce7b8f3e	//rax1 v30.16b,v25.16b,v27.16b			// D[1]
603
.inst	0xce7c8f5f	//rax1 v31.16b,v26.16b,v28.16b			// D[2]
604
.inst	0xce7d8f7b	//rax1 v27.16b,v27.16b,v29.16b			// D[3]
605
.inst	0xce798f9c	//rax1 v28.16b,v28.16b,v25.16b			// D[4]
606
.inst	0xce7a8fbd	//rax1 v29.16b,v29.16b,v26.16b			// D[0]
607

608
	////////////////////////////////////////////////// Theta+Rho+Pi
609
.inst	0xce9efc39	//xar v25.16b,   v1.16b,v30.16b,#64-1 // C[0]=A[2][0]
610

611
.inst	0xce9e50c1	//xar v1.16b,v6.16b,v30.16b,#64-44
612
.inst	0xce9cb126	//xar v6.16b,v9.16b,v28.16b,#64-20
613
.inst	0xce9f0ec9	//xar v9.16b,v22.16b,v31.16b,#64-61
614
.inst	0xce9c65d6	//xar v22.16b,v14.16b,v28.16b,#64-39
615
.inst	0xce9dba8e	//xar v14.16b,v20.16b,v29.16b,#64-18
616

617
.inst	0xce9f085a	//xar v26.16b,   v2.16b,v31.16b,#64-62 // C[1]=A[4][0]
618

619
.inst	0xce9f5582	//xar v2.16b,v12.16b,v31.16b,#64-43
620
.inst	0xce9b9dac	//xar v12.16b,v13.16b,v27.16b,#64-25
621
.inst	0xce9ce26d	//xar v13.16b,v19.16b,v28.16b,#64-8
622
.inst	0xce9b22f3	//xar v19.16b,v23.16b,v27.16b,#64-56
623
.inst	0xce9d5df7	//xar v23.16b,v15.16b,v29.16b,#64-41
624

625
.inst	0xce9c948f	//xar v15.16b,v4.16b,v28.16b,#64-27
626

627
.inst	0xce9ccb1c	//xar v28.16b,   v24.16b,v28.16b,#64-14 // D[4]=A[0][4]
628
.inst	0xce9efab8	//xar v24.16b,v21.16b,v30.16b,#64-2
629
.inst	0xce9b2508	//xar v8.16b,v8.16b,v27.16b,#64-55 // A[1][3]=A[4][1]
630
.inst	0xce9e4e04	//xar v4.16b,v16.16b,v30.16b,#64-45 // A[0][4]=A[1][3]
631
.inst	0xce9d70b0	//xar v16.16b,v5.16b,v29.16b,#64-36
632

633
.inst	0xce9b9065	//xar v5.16b,v3.16b,v27.16b,#64-28
634

635
	eor	v0.16b,v0.16b,v29.16b
636

637
.inst	0xce9bae5b	//xar v27.16b,   v18.16b,v27.16b,#64-21 // D[3]=A[0][3]
638
.inst	0xce9fc623	//xar v3.16b,v17.16b,v31.16b,#64-15 // A[0][3]=A[3][3]
639
.inst	0xce9ed97e	//xar v30.16b,   v11.16b,v30.16b,#64-10 // D[1]=A[3][2]
640
.inst	0xce9fe8ff	//xar v31.16b,   v7.16b,v31.16b,#64-6 // D[2]=A[2][1]
641
.inst	0xce9df55d	//xar v29.16b,   v10.16b,v29.16b,#64-3 // D[0]=A[1][2]
642

643
	////////////////////////////////////////////////// Chi+Iota
644
.inst	0xce362354	//bcax v20.16b,v26.16b,   v22.16b,v8.16b	// A[1][3]=A[4][1]
645
.inst	0xce375915	//bcax v21.16b,v8.16b,v23.16b,v22.16b	// A[1][3]=A[4][1]
646
.inst	0xce385ed6	//bcax v22.16b,v22.16b,v24.16b,v23.16b
647
.inst	0xce3a62f7	//bcax v23.16b,v23.16b,v26.16b,   v24.16b
648
.inst	0xce286b18	//bcax v24.16b,v24.16b,v8.16b,v26.16b	// A[1][3]=A[4][1]
649

650
	ld1r	{v26.2d},[x10],#8
651

652
.inst	0xce330fd1	//bcax v17.16b,v30.16b,   v19.16b,v3.16b	// A[0][3]=A[3][3]
653
.inst	0xce2f4c72	//bcax v18.16b,v3.16b,v15.16b,v19.16b	// A[0][3]=A[3][3]
654
.inst	0xce303e73	//bcax v19.16b,v19.16b,v16.16b,v15.16b
655
.inst	0xce3e41ef	//bcax v15.16b,v15.16b,v30.16b,   v16.16b
656
.inst	0xce237a10	//bcax v16.16b,v16.16b,v3.16b,v30.16b	// A[0][3]=A[3][3]
657

658
.inst	0xce2c7f2a	//bcax v10.16b,v25.16b,   v12.16b,v31.16b
659
.inst	0xce2d33eb	//bcax v11.16b,v31.16b,   v13.16b,v12.16b
660
.inst	0xce2e358c	//bcax v12.16b,v12.16b,v14.16b,v13.16b
661
.inst	0xce3939ad	//bcax v13.16b,v13.16b,v25.16b,   v14.16b
662
.inst	0xce3f65ce	//bcax v14.16b,v14.16b,v31.16b,   v25.16b
663

664
.inst	0xce2913a7	//bcax v7.16b,v29.16b,   v9.16b,v4.16b	// A[0][4]=A[1][3]
665
.inst	0xce252488	//bcax v8.16b,v4.16b,v5.16b,v9.16b	// A[0][4]=A[1][3]
666
.inst	0xce261529	//bcax v9.16b,v9.16b,v6.16b,v5.16b
667
.inst	0xce3d18a5	//bcax v5.16b,v5.16b,v29.16b,   v6.16b
668
.inst	0xce2474c6	//bcax v6.16b,v6.16b,v4.16b,v29.16b	// A[0][4]=A[1][3]
669

670
.inst	0xce207363	//bcax v3.16b,v27.16b,   v0.16b,v28.16b
671
.inst	0xce210384	//bcax v4.16b,v28.16b,   v1.16b,v0.16b
672
.inst	0xce220400	//bcax v0.16b,v0.16b,v2.16b,v1.16b
673
.inst	0xce3b0821	//bcax v1.16b,v1.16b,v27.16b,   v2.16b
674
.inst	0xce3c6c42	//bcax v2.16b,v2.16b,v28.16b,   v27.16b
675

676
	eor	v0.16b,v0.16b,v26.16b
677

678
	subs	x9,x9,#1
679
	bne	.Loop_ce
680

681
	ret
682
.size	KeccakF1600_ce,.-KeccakF1600_ce
683

684
.type	KeccakF1600_cext,%function
685
.align	5
686
KeccakF1600_cext:
687
	AARCH64_SIGN_LINK_REGISTER
688
	stp	x29,x30,[sp,#-80]!
689
	add	x29,sp,#0
690
	stp	d8,d9,[sp,#16]		// per ABI requirement
691
	stp	d10,d11,[sp,#32]
692
	stp	d12,d13,[sp,#48]
693
	stp	d14,d15,[sp,#64]
694
	ldp	d0,d1,[x0,#8*0]
695
	ldp	d2,d3,[x0,#8*2]
696
	ldp	d4,d5,[x0,#8*4]
697
	ldp	d6,d7,[x0,#8*6]
698
	ldp	d8,d9,[x0,#8*8]
699
	ldp	d10,d11,[x0,#8*10]
700
	ldp	d12,d13,[x0,#8*12]
701
	ldp	d14,d15,[x0,#8*14]
702
	ldp	d16,d17,[x0,#8*16]
703
	ldp	d18,d19,[x0,#8*18]
704
	ldp	d20,d21,[x0,#8*20]
705
	ldp	d22,d23,[x0,#8*22]
706
	ldr	d24,[x0,#8*24]
707
	bl	KeccakF1600_ce
708
	ldr	x30,[sp,#8]
709
	stp	d0,d1,[x0,#8*0]
710
	stp	d2,d3,[x0,#8*2]
711
	stp	d4,d5,[x0,#8*4]
712
	stp	d6,d7,[x0,#8*6]
713
	stp	d8,d9,[x0,#8*8]
714
	stp	d10,d11,[x0,#8*10]
715
	stp	d12,d13,[x0,#8*12]
716
	stp	d14,d15,[x0,#8*14]
717
	stp	d16,d17,[x0,#8*16]
718
	stp	d18,d19,[x0,#8*18]
719
	stp	d20,d21,[x0,#8*20]
720
	stp	d22,d23,[x0,#8*22]
721
	str	d24,[x0,#8*24]
722

723
	ldp	d8,d9,[sp,#16]
724
	ldp	d10,d11,[sp,#32]
725
	ldp	d12,d13,[sp,#48]
726
	ldp	d14,d15,[sp,#64]
727
	ldr	x29,[sp],#80
728
	AARCH64_VALIDATE_LINK_REGISTER
729
	ret
730
.size	KeccakF1600_cext,.-KeccakF1600_cext
731
.globl	SHA3_absorb_cext
732
.type	SHA3_absorb_cext,%function
733
.align	5
734
SHA3_absorb_cext:
735
	AARCH64_SIGN_LINK_REGISTER
736
	stp	x29,x30,[sp,#-80]!
737
	add	x29,sp,#0
738
	stp	d8,d9,[sp,#16]		// per ABI requirement
739
	stp	d10,d11,[sp,#32]
740
	stp	d12,d13,[sp,#48]
741
	stp	d14,d15,[sp,#64]
742
	ldp	d0,d1,[x0,#8*0]
743
	ldp	d2,d3,[x0,#8*2]
744
	ldp	d4,d5,[x0,#8*4]
745
	ldp	d6,d7,[x0,#8*6]
746
	ldp	d8,d9,[x0,#8*8]
747
	ldp	d10,d11,[x0,#8*10]
748
	ldp	d12,d13,[x0,#8*12]
749
	ldp	d14,d15,[x0,#8*14]
750
	ldp	d16,d17,[x0,#8*16]
751
	ldp	d18,d19,[x0,#8*18]
752
	ldp	d20,d21,[x0,#8*20]
753
	ldp	d22,d23,[x0,#8*22]
754
	ldr	d24,[x0,#8*24]
755
	b	.Loop_absorb_ce
756

757
.align	4
758
.Loop_absorb_ce:
759
	subs	x2,x2,x3		// len - bsz
760
	blo	.Labsorbed_ce
761
	ldr	d31,[x1],#8		// *inp++
762
#ifdef	__AARCH64EB__
763
	rev64	v31.16b,v31.16b
764
#endif
765
	eor	v0.16b,v0.16b,v31.16b
766
	cmp	x3,#8*(0+2)
767
	blo	.Lprocess_block_ce
768
	ldr	d31,[x1],#8		// *inp++
769
#ifdef	__AARCH64EB__
770
	rev64	v31.16b,v31.16b
771
#endif
772
	eor	v1.16b,v1.16b,v31.16b
773
	beq	.Lprocess_block_ce
774
	ldr	d31,[x1],#8		// *inp++
775
#ifdef	__AARCH64EB__
776
	rev64	v31.16b,v31.16b
777
#endif
778
	eor	v2.16b,v2.16b,v31.16b
779
	cmp	x3,#8*(2+2)
780
	blo	.Lprocess_block_ce
781
	ldr	d31,[x1],#8		// *inp++
782
#ifdef	__AARCH64EB__
783
	rev64	v31.16b,v31.16b
784
#endif
785
	eor	v3.16b,v3.16b,v31.16b
786
	beq	.Lprocess_block_ce
787
	ldr	d31,[x1],#8		// *inp++
788
#ifdef	__AARCH64EB__
789
	rev64	v31.16b,v31.16b
790
#endif
791
	eor	v4.16b,v4.16b,v31.16b
792
	cmp	x3,#8*(4+2)
793
	blo	.Lprocess_block_ce
794
	ldr	d31,[x1],#8		// *inp++
795
#ifdef	__AARCH64EB__
796
	rev64	v31.16b,v31.16b
797
#endif
798
	eor	v5.16b,v5.16b,v31.16b
799
	beq	.Lprocess_block_ce
800
	ldr	d31,[x1],#8		// *inp++
801
#ifdef	__AARCH64EB__
802
	rev64	v31.16b,v31.16b
803
#endif
804
	eor	v6.16b,v6.16b,v31.16b
805
	cmp	x3,#8*(6+2)
806
	blo	.Lprocess_block_ce
807
	ldr	d31,[x1],#8		// *inp++
808
#ifdef	__AARCH64EB__
809
	rev64	v31.16b,v31.16b
810
#endif
811
	eor	v7.16b,v7.16b,v31.16b
812
	beq	.Lprocess_block_ce
813
	ldr	d31,[x1],#8		// *inp++
814
#ifdef	__AARCH64EB__
815
	rev64	v31.16b,v31.16b
816
#endif
817
	eor	v8.16b,v8.16b,v31.16b
818
	cmp	x3,#8*(8+2)
819
	blo	.Lprocess_block_ce
820
	ldr	d31,[x1],#8		// *inp++
821
#ifdef	__AARCH64EB__
822
	rev64	v31.16b,v31.16b
823
#endif
824
	eor	v9.16b,v9.16b,v31.16b
825
	beq	.Lprocess_block_ce
826
	ldr	d31,[x1],#8		// *inp++
827
#ifdef	__AARCH64EB__
828
	rev64	v31.16b,v31.16b
829
#endif
830
	eor	v10.16b,v10.16b,v31.16b
831
	cmp	x3,#8*(10+2)
832
	blo	.Lprocess_block_ce
833
	ldr	d31,[x1],#8		// *inp++
834
#ifdef	__AARCH64EB__
835
	rev64	v31.16b,v31.16b
836
#endif
837
	eor	v11.16b,v11.16b,v31.16b
838
	beq	.Lprocess_block_ce
839
	ldr	d31,[x1],#8		// *inp++
840
#ifdef	__AARCH64EB__
841
	rev64	v31.16b,v31.16b
842
#endif
843
	eor	v12.16b,v12.16b,v31.16b
844
	cmp	x3,#8*(12+2)
845
	blo	.Lprocess_block_ce
846
	ldr	d31,[x1],#8		// *inp++
847
#ifdef	__AARCH64EB__
848
	rev64	v31.16b,v31.16b
849
#endif
850
	eor	v13.16b,v13.16b,v31.16b
851
	beq	.Lprocess_block_ce
852
	ldr	d31,[x1],#8		// *inp++
853
#ifdef	__AARCH64EB__
854
	rev64	v31.16b,v31.16b
855
#endif
856
	eor	v14.16b,v14.16b,v31.16b
857
	cmp	x3,#8*(14+2)
858
	blo	.Lprocess_block_ce
859
	ldr	d31,[x1],#8		// *inp++
860
#ifdef	__AARCH64EB__
861
	rev64	v31.16b,v31.16b
862
#endif
863
	eor	v15.16b,v15.16b,v31.16b
864
	beq	.Lprocess_block_ce
865
	ldr	d31,[x1],#8		// *inp++
866
#ifdef	__AARCH64EB__
867
	rev64	v31.16b,v31.16b
868
#endif
869
	eor	v16.16b,v16.16b,v31.16b
870
	cmp	x3,#8*(16+2)
871
	blo	.Lprocess_block_ce
872
	ldr	d31,[x1],#8		// *inp++
873
#ifdef	__AARCH64EB__
874
	rev64	v31.16b,v31.16b
875
#endif
876
	eor	v17.16b,v17.16b,v31.16b
877
	beq	.Lprocess_block_ce
878
	ldr	d31,[x1],#8		// *inp++
879
#ifdef	__AARCH64EB__
880
	rev64	v31.16b,v31.16b
881
#endif
882
	eor	v18.16b,v18.16b,v31.16b
883
	cmp	x3,#8*(18+2)
884
	blo	.Lprocess_block_ce
885
	ldr	d31,[x1],#8		// *inp++
886
#ifdef	__AARCH64EB__
887
	rev64	v31.16b,v31.16b
888
#endif
889
	eor	v19.16b,v19.16b,v31.16b
890
	beq	.Lprocess_block_ce
891
	ldr	d31,[x1],#8		// *inp++
892
#ifdef	__AARCH64EB__
893
	rev64	v31.16b,v31.16b
894
#endif
895
	eor	v20.16b,v20.16b,v31.16b
896
	cmp	x3,#8*(20+2)
897
	blo	.Lprocess_block_ce
898
	ldr	d31,[x1],#8		// *inp++
899
#ifdef	__AARCH64EB__
900
	rev64	v31.16b,v31.16b
901
#endif
902
	eor	v21.16b,v21.16b,v31.16b
903
	beq	.Lprocess_block_ce
904
	ldr	d31,[x1],#8		// *inp++
905
#ifdef	__AARCH64EB__
906
	rev64	v31.16b,v31.16b
907
#endif
908
	eor	v22.16b,v22.16b,v31.16b
909
	cmp	x3,#8*(22+2)
910
	blo	.Lprocess_block_ce
911
	ldr	d31,[x1],#8		// *inp++
912
#ifdef	__AARCH64EB__
913
	rev64	v31.16b,v31.16b
914
#endif
915
	eor	v23.16b,v23.16b,v31.16b
916
	beq	.Lprocess_block_ce
917
	ldr	d31,[x1],#8		// *inp++
918
#ifdef	__AARCH64EB__
919
	rev64	v31.16b,v31.16b
920
#endif
921
	eor	v24.16b,v24.16b,v31.16b
922

923
.Lprocess_block_ce:
924

925
	bl	KeccakF1600_ce
926

927
	b	.Loop_absorb_ce
928

929
.align	4
930
.Labsorbed_ce:
931
	stp	d0,d1,[x0,#8*0]
932
	stp	d2,d3,[x0,#8*2]
933
	stp	d4,d5,[x0,#8*4]
934
	stp	d6,d7,[x0,#8*6]
935
	stp	d8,d9,[x0,#8*8]
936
	stp	d10,d11,[x0,#8*10]
937
	stp	d12,d13,[x0,#8*12]
938
	stp	d14,d15,[x0,#8*14]
939
	stp	d16,d17,[x0,#8*16]
940
	stp	d18,d19,[x0,#8*18]
941
	stp	d20,d21,[x0,#8*20]
942
	stp	d22,d23,[x0,#8*22]
943
	str	d24,[x0,#8*24]
944
	add	x0,x2,x3		// return value
945

946
	ldp	d8,d9,[sp,#16]
947
	ldp	d10,d11,[sp,#32]
948
	ldp	d12,d13,[sp,#48]
949
	ldp	d14,d15,[sp,#64]
950
	ldp	x29,x30,[sp],#80
951
	AARCH64_VALIDATE_LINK_REGISTER
952
	ret
953
.size	SHA3_absorb_cext,.-SHA3_absorb_cext
954
.globl	SHA3_squeeze_cext
955
.type	SHA3_squeeze_cext,%function
956
.align	5
957
SHA3_squeeze_cext:
958
	AARCH64_SIGN_LINK_REGISTER
959
	stp	x29,x30,[sp,#-16]!
960
	add	x29,sp,#0
961
	mov	x9,x0
962
	mov	x10,x3
963

964
.Loop_squeeze_ce:
965
	ldr	x4,[x9],#8
966
	cmp	x2,#8
967
	blo	.Lsqueeze_tail_ce
968
#ifdef	__AARCH64EB__
969
	rev	x4,x4
970
#endif
971
	str	x4,[x1],#8
972
	beq	.Lsqueeze_done_ce
973

974
	sub	x2,x2,#8
975
	subs	x10,x10,#8
976
	bhi	.Loop_squeeze_ce
977

978
	bl	KeccakF1600_cext
979
	ldr	x30,[sp,#8]
980
	mov	x9,x0
981
	mov	x10,x3
982
	b	.Loop_squeeze_ce
983

984
.align	4
985
.Lsqueeze_tail_ce:
986
	strb	w4,[x1],#1
987
	lsr	x4,x4,#8
988
	subs	x2,x2,#1
989
	beq	.Lsqueeze_done_ce
990
	strb	w4,[x1],#1
991
	lsr	x4,x4,#8
992
	subs	x2,x2,#1
993
	beq	.Lsqueeze_done_ce
994
	strb	w4,[x1],#1
995
	lsr	x4,x4,#8
996
	subs	x2,x2,#1
997
	beq	.Lsqueeze_done_ce
998
	strb	w4,[x1],#1
999
	lsr	x4,x4,#8
1000
	subs	x2,x2,#1
1001
	beq	.Lsqueeze_done_ce
1002
	strb	w4,[x1],#1
1003
	lsr	x4,x4,#8
1004
	subs	x2,x2,#1
1005
	beq	.Lsqueeze_done_ce
1006
	strb	w4,[x1],#1
1007
	lsr	x4,x4,#8
1008
	subs	x2,x2,#1
1009
	beq	.Lsqueeze_done_ce
1010
	strb	w4,[x1],#1
1011

1012
.Lsqueeze_done_ce:
1013
	ldr	x29,[sp],#16
1014
	AARCH64_VALIDATE_LINK_REGISTER
1015
	ret
1016
.size	SHA3_squeeze_cext,.-SHA3_squeeze_cext
1017
.byte	75,101,99,99,97,107,45,49,54,48,48,32,97,98,115,111,114,98,32,97,110,100,32,115,113,117,101,101,122,101,32,102,111,114,32,65,82,77,118,56,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
1018
.align	2
1019

1020
Product

Resources

Company
