1/* $FreeBSD$ */
2/* Do not modify. This file is auto-generated from keccak1600-armv8.pl. */
3.text
4
5.align	8	// strategic alignment and padding that allows to use
6		// address value as loop termination condition...
7.quad	0,0,0,0,0,0,0,0
8.type	iotas,%object
9iotas:
10.quad	0x0000000000000001
11.quad	0x0000000000008082
12.quad	0x800000000000808a
13.quad	0x8000000080008000
14.quad	0x000000000000808b
15.quad	0x0000000080000001
16.quad	0x8000000080008081
17.quad	0x8000000000008009
18.quad	0x000000000000008a
19.quad	0x0000000000000088
20.quad	0x0000000080008009
21.quad	0x000000008000000a
22.quad	0x000000008000808b
23.quad	0x800000000000008b
24.quad	0x8000000000008089
25.quad	0x8000000000008003
26.quad	0x8000000000008002
27.quad	0x8000000000000080
28.quad	0x000000000000800a
29.quad	0x800000008000000a
30.quad	0x8000000080008081
31.quad	0x8000000000008080
32.quad	0x0000000080000001
33.quad	0x8000000080008008
34.size	iotas,.-iotas
35.type	KeccakF1600_int,%function
36.align	5
37KeccakF1600_int:
38	adr	x28,iotas
39.inst	0xd503233f			// paciasp
40	stp	x28,x30,[sp,#16]		// 32 bytes on top are mine
41	b	.Loop
42.align	4
43.Loop:
44	////////////////////////////////////////// Theta
45	eor	x26,x0,x5
46	stp	x4,x9,[sp,#0]	// offload pair...
47	eor	x27,x1,x6
48	eor	x28,x2,x7
49	eor	x30,x3,x8
50	eor	x4,x4,x9
51	eor	x26,x26,x10
52	eor	x27,x27,x11
53	eor	x28,x28,x12
54	eor	x30,x30,x13
55	eor	x4,x4,x14
56	eor	x26,x26,x15
57	eor	x27,x27,x16
58	eor	x28,x28,x17
59	eor	x30,x30,x25
60	eor	x4,x4,x19
61	eor	x26,x26,x20
62	eor	x28,x28,x22
63	eor	x27,x27,x21
64	eor	x30,x30,x23
65	eor	x4,x4,x24
66
67	eor	x9,x26,x28,ror#63
68
69	eor	x1,x1,x9
70	eor	x6,x6,x9
71	eor	x11,x11,x9
72	eor	x16,x16,x9
73	eor	x21,x21,x9
74
75	eor	x9,x27,x30,ror#63
76	eor	x28,x28,x4,ror#63
77	eor	x30,x30,x26,ror#63
78	eor	x4,x4,x27,ror#63
79
80	eor	x27,   x2,x9		// mov	x27,x2
81	eor	x7,x7,x9
82	eor	x12,x12,x9
83	eor	x17,x17,x9
84	eor	x22,x22,x9
85
86	eor	x0,x0,x4
87	eor	x5,x5,x4
88	eor	x10,x10,x4
89	eor	x15,x15,x4
90	eor	x20,x20,x4
91	ldp	x4,x9,[sp,#0]	// re-load offloaded data
92	eor	x26,   x3,x28		// mov	x26,x3
93	eor	x8,x8,x28
94	eor	x13,x13,x28
95	eor	x25,x25,x28
96	eor	x23,x23,x28
97
98	eor	x28,   x4,x30		// mov	x28,x4
99	eor	x9,x9,x30
100	eor	x14,x14,x30
101	eor	x19,x19,x30
102	eor	x24,x24,x30
103
104	////////////////////////////////////////// Rho+Pi
105	mov	x30,x1
106	ror	x1,x6,#64-44
107	//mov	x27,x2
108	ror	x2,x12,#64-43
109	//mov	x26,x3
110	ror	x3,x25,#64-21
111	//mov	x28,x4
112	ror	x4,x24,#64-14
113
114	ror	x6,x9,#64-20
115	ror	x12,x13,#64-25
116	ror	x25,x17,#64-15
117	ror	x24,x21,#64-2
118
119	ror	x9,x22,#64-61
120	ror	x13,x19,#64-8
121	ror	x17,x11,#64-10
122	ror	x21,x8,#64-55
123
124	ror	x22,x14,#64-39
125	ror	x19,x23,#64-56
126	ror	x11,x7,#64-6
127	ror	x8,x16,#64-45
128
129	ror	x14,x20,#64-18
130	ror	x23,x15,#64-41
131	ror	x7,x10,#64-3
132	ror	x16,x5,#64-36
133
134	ror	x5,x26,#64-28
135	ror	x10,x30,#64-1
136	ror	x15,x28,#64-27
137	ror	x20,x27,#64-62
138
139	////////////////////////////////////////// Chi+Iota
140	bic	x26,x2,x1
141	bic	x27,x3,x2
142	bic	x28,x0,x4
143	bic	x30,x1,x0
144	eor	x0,x0,x26
145	bic	x26,x4,x3
146	eor	x1,x1,x27
147	ldr	x27,[sp,#16]
148	eor	x3,x3,x28
149	eor	x4,x4,x30
150	eor	x2,x2,x26
151	ldr	x30,[x27],#8		// Iota[i++]
152
153	bic	x26,x7,x6
154	tst	x27,#255			// are we done?
155	str	x27,[sp,#16]
156	bic	x27,x8,x7
157	bic	x28,x5,x9
158	eor	x0,x0,x30		// A[0][0] ^= Iota
159	bic	x30,x6,x5
160	eor	x5,x5,x26
161	bic	x26,x9,x8
162	eor	x6,x6,x27
163	eor	x8,x8,x28
164	eor	x9,x9,x30
165	eor	x7,x7,x26
166
167	bic	x26,x12,x11
168	bic	x27,x13,x12
169	bic	x28,x10,x14
170	bic	x30,x11,x10
171	eor	x10,x10,x26
172	bic	x26,x14,x13
173	eor	x11,x11,x27
174	eor	x13,x13,x28
175	eor	x14,x14,x30
176	eor	x12,x12,x26
177
178	bic	x26,x17,x16
179	bic	x27,x25,x17
180	bic	x28,x15,x19
181	bic	x30,x16,x15
182	eor	x15,x15,x26
183	bic	x26,x19,x25
184	eor	x16,x16,x27
185	eor	x25,x25,x28
186	eor	x19,x19,x30
187	eor	x17,x17,x26
188
189	bic	x26,x22,x21
190	bic	x27,x23,x22
191	bic	x28,x20,x24
192	bic	x30,x21,x20
193	eor	x20,x20,x26
194	bic	x26,x24,x23
195	eor	x21,x21,x27
196	eor	x23,x23,x28
197	eor	x24,x24,x30
198	eor	x22,x22,x26
199
200	bne	.Loop
201
202	ldr	x30,[sp,#24]
203.inst	0xd50323bf			// autiasp
204	ret
205.size	KeccakF1600_int,.-KeccakF1600_int
206
207.type	KeccakF1600,%function
208.align	5
209KeccakF1600:
210.inst	0xd503233f			// paciasp
211	stp	x29,x30,[sp,#-128]!
212	add	x29,sp,#0
213	stp	x19,x20,[sp,#16]
214	stp	x21,x22,[sp,#32]
215	stp	x23,x24,[sp,#48]
216	stp	x25,x26,[sp,#64]
217	stp	x27,x28,[sp,#80]
218	sub	sp,sp,#48
219
220	str	x0,[sp,#32]			// offload argument
221	mov	x26,x0
222	ldp	x0,x1,[x0,#16*0]
223	ldp	x2,x3,[x26,#16*1]
224	ldp	x4,x5,[x26,#16*2]
225	ldp	x6,x7,[x26,#16*3]
226	ldp	x8,x9,[x26,#16*4]
227	ldp	x10,x11,[x26,#16*5]
228	ldp	x12,x13,[x26,#16*6]
229	ldp	x14,x15,[x26,#16*7]
230	ldp	x16,x17,[x26,#16*8]
231	ldp	x25,x19,[x26,#16*9]
232	ldp	x20,x21,[x26,#16*10]
233	ldp	x22,x23,[x26,#16*11]
234	ldr	x24,[x26,#16*12]
235
236	bl	KeccakF1600_int
237
238	ldr	x26,[sp,#32]
239	stp	x0,x1,[x26,#16*0]
240	stp	x2,x3,[x26,#16*1]
241	stp	x4,x5,[x26,#16*2]
242	stp	x6,x7,[x26,#16*3]
243	stp	x8,x9,[x26,#16*4]
244	stp	x10,x11,[x26,#16*5]
245	stp	x12,x13,[x26,#16*6]
246	stp	x14,x15,[x26,#16*7]
247	stp	x16,x17,[x26,#16*8]
248	stp	x25,x19,[x26,#16*9]
249	stp	x20,x21,[x26,#16*10]
250	stp	x22,x23,[x26,#16*11]
251	str	x24,[x26,#16*12]
252
253	ldp	x19,x20,[x29,#16]
254	add	sp,sp,#48
255	ldp	x21,x22,[x29,#32]
256	ldp	x23,x24,[x29,#48]
257	ldp	x25,x26,[x29,#64]
258	ldp	x27,x28,[x29,#80]
259	ldp	x29,x30,[sp],#128
260.inst	0xd50323bf			// autiasp
261	ret
262.size	KeccakF1600,.-KeccakF1600
263
264.globl	SHA3_absorb
265.type	SHA3_absorb,%function
266.align	5
267SHA3_absorb:
268.inst	0xd503233f			// paciasp
269	stp	x29,x30,[sp,#-128]!
270	add	x29,sp,#0
271	stp	x19,x20,[sp,#16]
272	stp	x21,x22,[sp,#32]
273	stp	x23,x24,[sp,#48]
274	stp	x25,x26,[sp,#64]
275	stp	x27,x28,[sp,#80]
276	sub	sp,sp,#64
277
278	stp	x0,x1,[sp,#32]			// offload arguments
279	stp	x2,x3,[sp,#48]
280
281	mov	x26,x0			// uint64_t A[5][5]
282	mov	x27,x1			// const void *inp
283	mov	x28,x2			// size_t len
284	mov	x30,x3			// size_t bsz
285	ldp	x0,x1,[x26,#16*0]
286	ldp	x2,x3,[x26,#16*1]
287	ldp	x4,x5,[x26,#16*2]
288	ldp	x6,x7,[x26,#16*3]
289	ldp	x8,x9,[x26,#16*4]
290	ldp	x10,x11,[x26,#16*5]
291	ldp	x12,x13,[x26,#16*6]
292	ldp	x14,x15,[x26,#16*7]
293	ldp	x16,x17,[x26,#16*8]
294	ldp	x25,x19,[x26,#16*9]
295	ldp	x20,x21,[x26,#16*10]
296	ldp	x22,x23,[x26,#16*11]
297	ldr	x24,[x26,#16*12]
298	b	.Loop_absorb
299
300.align	4
301.Loop_absorb:
302	subs	x26,x28,x30		// len - bsz
303	blo	.Labsorbed
304
305	str	x26,[sp,#48]			// save len - bsz
306	ldr	x26,[x27],#8		// *inp++
307#ifdef	__AARCH64EB__
308	rev	x26,x26
309#endif
310	eor	x0,x0,x26
311	cmp	x30,#8*(0+2)
312	blo	.Lprocess_block
313	ldr	x26,[x27],#8		// *inp++
314#ifdef	__AARCH64EB__
315	rev	x26,x26
316#endif
317	eor	x1,x1,x26
318	beq	.Lprocess_block
319	ldr	x26,[x27],#8		// *inp++
320#ifdef	__AARCH64EB__
321	rev	x26,x26
322#endif
323	eor	x2,x2,x26
324	cmp	x30,#8*(2+2)
325	blo	.Lprocess_block
326	ldr	x26,[x27],#8		// *inp++
327#ifdef	__AARCH64EB__
328	rev	x26,x26
329#endif
330	eor	x3,x3,x26
331	beq	.Lprocess_block
332	ldr	x26,[x27],#8		// *inp++
333#ifdef	__AARCH64EB__
334	rev	x26,x26
335#endif
336	eor	x4,x4,x26
337	cmp	x30,#8*(4+2)
338	blo	.Lprocess_block
339	ldr	x26,[x27],#8		// *inp++
340#ifdef	__AARCH64EB__
341	rev	x26,x26
342#endif
343	eor	x5,x5,x26
344	beq	.Lprocess_block
345	ldr	x26,[x27],#8		// *inp++
346#ifdef	__AARCH64EB__
347	rev	x26,x26
348#endif
349	eor	x6,x6,x26
350	cmp	x30,#8*(6+2)
351	blo	.Lprocess_block
352	ldr	x26,[x27],#8		// *inp++
353#ifdef	__AARCH64EB__
354	rev	x26,x26
355#endif
356	eor	x7,x7,x26
357	beq	.Lprocess_block
358	ldr	x26,[x27],#8		// *inp++
359#ifdef	__AARCH64EB__
360	rev	x26,x26
361#endif
362	eor	x8,x8,x26
363	cmp	x30,#8*(8+2)
364	blo	.Lprocess_block
365	ldr	x26,[x27],#8		// *inp++
366#ifdef	__AARCH64EB__
367	rev	x26,x26
368#endif
369	eor	x9,x9,x26
370	beq	.Lprocess_block
371	ldr	x26,[x27],#8		// *inp++
372#ifdef	__AARCH64EB__
373	rev	x26,x26
374#endif
375	eor	x10,x10,x26
376	cmp	x30,#8*(10+2)
377	blo	.Lprocess_block
378	ldr	x26,[x27],#8		// *inp++
379#ifdef	__AARCH64EB__
380	rev	x26,x26
381#endif
382	eor	x11,x11,x26
383	beq	.Lprocess_block
384	ldr	x26,[x27],#8		// *inp++
385#ifdef	__AARCH64EB__
386	rev	x26,x26
387#endif
388	eor	x12,x12,x26
389	cmp	x30,#8*(12+2)
390	blo	.Lprocess_block
391	ldr	x26,[x27],#8		// *inp++
392#ifdef	__AARCH64EB__
393	rev	x26,x26
394#endif
395	eor	x13,x13,x26
396	beq	.Lprocess_block
397	ldr	x26,[x27],#8		// *inp++
398#ifdef	__AARCH64EB__
399	rev	x26,x26
400#endif
401	eor	x14,x14,x26
402	cmp	x30,#8*(14+2)
403	blo	.Lprocess_block
404	ldr	x26,[x27],#8		// *inp++
405#ifdef	__AARCH64EB__
406	rev	x26,x26
407#endif
408	eor	x15,x15,x26
409	beq	.Lprocess_block
410	ldr	x26,[x27],#8		// *inp++
411#ifdef	__AARCH64EB__
412	rev	x26,x26
413#endif
414	eor	x16,x16,x26
415	cmp	x30,#8*(16+2)
416	blo	.Lprocess_block
417	ldr	x26,[x27],#8		// *inp++
418#ifdef	__AARCH64EB__
419	rev	x26,x26
420#endif
421	eor	x17,x17,x26
422	beq	.Lprocess_block
423	ldr	x26,[x27],#8		// *inp++
424#ifdef	__AARCH64EB__
425	rev	x26,x26
426#endif
427	eor	x25,x25,x26
428	cmp	x30,#8*(18+2)
429	blo	.Lprocess_block
430	ldr	x26,[x27],#8		// *inp++
431#ifdef	__AARCH64EB__
432	rev	x26,x26
433#endif
434	eor	x19,x19,x26
435	beq	.Lprocess_block
436	ldr	x26,[x27],#8		// *inp++
437#ifdef	__AARCH64EB__
438	rev	x26,x26
439#endif
440	eor	x20,x20,x26
441	cmp	x30,#8*(20+2)
442	blo	.Lprocess_block
443	ldr	x26,[x27],#8		// *inp++
444#ifdef	__AARCH64EB__
445	rev	x26,x26
446#endif
447	eor	x21,x21,x26
448	beq	.Lprocess_block
449	ldr	x26,[x27],#8		// *inp++
450#ifdef	__AARCH64EB__
451	rev	x26,x26
452#endif
453	eor	x22,x22,x26
454	cmp	x30,#8*(22+2)
455	blo	.Lprocess_block
456	ldr	x26,[x27],#8		// *inp++
457#ifdef	__AARCH64EB__
458	rev	x26,x26
459#endif
460	eor	x23,x23,x26
461	beq	.Lprocess_block
462	ldr	x26,[x27],#8		// *inp++
463#ifdef	__AARCH64EB__
464	rev	x26,x26
465#endif
466	eor	x24,x24,x26
467
468.Lprocess_block:
469	str	x27,[sp,#40]			// save inp
470
471	bl	KeccakF1600_int
472
473	ldr	x27,[sp,#40]			// restore arguments
474	ldp	x28,x30,[sp,#48]
475	b	.Loop_absorb
476
477.align	4
478.Labsorbed:
479	ldr	x27,[sp,#32]
480	stp	x0,x1,[x27,#16*0]
481	stp	x2,x3,[x27,#16*1]
482	stp	x4,x5,[x27,#16*2]
483	stp	x6,x7,[x27,#16*3]
484	stp	x8,x9,[x27,#16*4]
485	stp	x10,x11,[x27,#16*5]
486	stp	x12,x13,[x27,#16*6]
487	stp	x14,x15,[x27,#16*7]
488	stp	x16,x17,[x27,#16*8]
489	stp	x25,x19,[x27,#16*9]
490	stp	x20,x21,[x27,#16*10]
491	stp	x22,x23,[x27,#16*11]
492	str	x24,[x27,#16*12]
493
494	mov	x0,x28			// return value
495	ldp	x19,x20,[x29,#16]
496	add	sp,sp,#64
497	ldp	x21,x22,[x29,#32]
498	ldp	x23,x24,[x29,#48]
499	ldp	x25,x26,[x29,#64]
500	ldp	x27,x28,[x29,#80]
501	ldp	x29,x30,[sp],#128
502.inst	0xd50323bf			// autiasp
503	ret
504.size	SHA3_absorb,.-SHA3_absorb
505.globl	SHA3_squeeze
506.type	SHA3_squeeze,%function
507.align	5
508SHA3_squeeze:
509.inst	0xd503233f			// paciasp
510	stp	x29,x30,[sp,#-48]!
511	add	x29,sp,#0
512	stp	x19,x20,[sp,#16]
513	stp	x21,x22,[sp,#32]
514
515	mov	x19,x0			// put aside arguments
516	mov	x20,x1
517	mov	x21,x2
518	mov	x22,x3
519
520.Loop_squeeze:
521	ldr	x4,[x0],#8
522	cmp	x21,#8
523	blo	.Lsqueeze_tail
524#ifdef	__AARCH64EB__
525	rev	x4,x4
526#endif
527	str	x4,[x20],#8
528	subs	x21,x21,#8
529	beq	.Lsqueeze_done
530
531	subs	x3,x3,#8
532	bhi	.Loop_squeeze
533
534	mov	x0,x19
535	bl	KeccakF1600
536	mov	x0,x19
537	mov	x3,x22
538	b	.Loop_squeeze
539
540.align	4
541.Lsqueeze_tail:
542	strb	w4,[x20],#1
543	lsr	x4,x4,#8
544	subs	x21,x21,#1
545	beq	.Lsqueeze_done
546	strb	w4,[x20],#1
547	lsr	x4,x4,#8
548	subs	x21,x21,#1
549	beq	.Lsqueeze_done
550	strb	w4,[x20],#1
551	lsr	x4,x4,#8
552	subs	x21,x21,#1
553	beq	.Lsqueeze_done
554	strb	w4,[x20],#1
555	lsr	x4,x4,#8
556	subs	x21,x21,#1
557	beq	.Lsqueeze_done
558	strb	w4,[x20],#1
559	lsr	x4,x4,#8
560	subs	x21,x21,#1
561	beq	.Lsqueeze_done
562	strb	w4,[x20],#1
563	lsr	x4,x4,#8
564	subs	x21,x21,#1
565	beq	.Lsqueeze_done
566	strb	w4,[x20],#1
567
568.Lsqueeze_done:
569	ldp	x19,x20,[sp,#16]
570	ldp	x21,x22,[sp,#32]
571	ldp	x29,x30,[sp],#48
572.inst	0xd50323bf			// autiasp
573	ret
574.size	SHA3_squeeze,.-SHA3_squeeze
575.type	KeccakF1600_ce,%function
576.align	5
577KeccakF1600_ce:
578	mov	x9,#12
579	adr	x10,iotas
580	b	.Loop_ce
581.align	4
582.Loop_ce:
583	////////////////////////////////////////////////// Theta
584.inst	0xce052819	//eor3 v25.16b,v0.16b,v5.16b,v10.16b
585.inst	0xce062c3a	//eor3 v26.16b,v1.16b,v6.16b,v11.16b
586.inst	0xce07305b	//eor3 v27.16b,v2.16b,v7.16b,v12.16b
587.inst	0xce08347c	//eor3 v28.16b,v3.16b,v8.16b,v13.16b
588.inst	0xce09389d	//eor3 v29.16b,v4.16b,v9.16b,v14.16b
589.inst	0xce0f5339	//eor3 v25.16b,v25.16b,   v15.16b,v20.16b
590.inst	0xce10575a	//eor3 v26.16b,v26.16b,   v16.16b,v21.16b
591.inst	0xce115b7b	//eor3 v27.16b,v27.16b,   v17.16b,v22.16b
592.inst	0xce125f9c	//eor3 v28.16b,v28.16b,   v18.16b,v23.16b
593.inst	0xce1363bd	//eor3 v29.16b,v29.16b,   v19.16b,v24.16b
594
595.inst	0xce7b8f3e	//rax1 v30.16b,v25.16b,v27.16b			// D[1]
596.inst	0xce7c8f5f	//rax1 v31.16b,v26.16b,v28.16b			// D[2]
597.inst	0xce7d8f7b	//rax1 v27.16b,v27.16b,v29.16b			// D[3]
598.inst	0xce798f9c	//rax1 v28.16b,v28.16b,v25.16b			// D[4]
599.inst	0xce7a8fbd	//rax1 v29.16b,v29.16b,v26.16b			// D[0]
600
601	////////////////////////////////////////////////// Theta+Rho+Pi
602.inst	0xce9e50d9	//xar v25.16b,   v6.16b,v30.16b,#64-44	// C[0]=A[0][1]
603.inst	0xce9cb126	//xar v6.16b,v9.16b,v28.16b,#64-20
604.inst	0xce9f0ec9	//xar v9.16b,v22.16b,v31.16b,#64-61
605.inst	0xce9c65d6	//xar v22.16b,v14.16b,v28.16b,#64-39
606.inst	0xce9dba8e	//xar v14.16b,v20.16b,v29.16b,#64-18
607
608.inst	0xce9f0854	//xar v20.16b,v2.16b,v31.16b,#64-62
609
610.inst	0xce9f5582	//xar v2.16b,v12.16b,v31.16b,#64-43
611.inst	0xce9b9dac	//xar v12.16b,v13.16b,v27.16b,#64-25
612.inst	0xce9ce26d	//xar v13.16b,v19.16b,v28.16b,#64-8
613.inst	0xce9b22f3	//xar v19.16b,v23.16b,v27.16b,#64-56
614.inst	0xce9d5df7	//xar v23.16b,v15.16b,v29.16b,#64-41
615
616.inst	0xce9c948f	//xar v15.16b,v4.16b,v28.16b,#64-27
617
618	eor	v0.16b,v0.16b,v29.16b
619	ldr	x11,[x10],#8
620
621.inst	0xce9bae5a	//xar v26.16b,   v18.16b,v27.16b,#64-21	// C[1]=A[0][3]
622.inst	0xce9fc632	//xar v18.16b,v17.16b,v31.16b,#64-15
623.inst	0xce9ed971	//xar v17.16b,v11.16b,v30.16b,#64-10
624.inst	0xce9fe8eb	//xar v11.16b,v7.16b,v31.16b,#64-6
625.inst	0xce9df547	//xar v7.16b,v10.16b,v29.16b,#64-3
626
627.inst	0xce9efc2a	//xar v10.16b,v1.16b,v30.16b,#64-1	// *
628
629.inst	0xce9ccb04	//xar v4.16b,v24.16b,v28.16b,#64-14
630.inst	0xce9efab8	//xar v24.16b,v21.16b,v30.16b,#64-2
631.inst	0xce9b2515	//xar v21.16b,v8.16b,v27.16b,#64-55
632.inst	0xce9e4e08	//xar v8.16b,v16.16b,v30.16b,#64-45
633.inst	0xce9d70b0	//xar v16.16b,v5.16b,v29.16b,#64-36
634
635.inst	0xce9b907b	//xar v27.16b,   v3.16b,v27.16b,#64-28	// C[2]=A[1][0]
636
637	////////////////////////////////////////////////// Chi+Iota
638	dup	v31.2d,x11				// borrow C[6]
639.inst	0xce22641c	//bcax v28.16b,   v0.16b,v2.16b,v25.16b	// *
640.inst	0xce3a0b21	//bcax v1.16b,v25.16b,   v26.16b,   v2.16b	// *
641.inst	0xce246842	//bcax v2.16b,v2.16b,v4.16b,v26.16b
642.inst	0xce201343	//bcax v3.16b,v26.16b,   v0.16b,v4.16b
643.inst	0xce390084	//bcax v4.16b,v4.16b,v25.16b,   v0.16b
644
645.inst	0xce271b65	//bcax v5.16b,v27.16b,   v7.16b,v6.16b	// *
646.inst	0xce281cd9	//bcax v25.16b,   v6.16b,v8.16b,v7.16b	// *
647.inst	0xce2920e7	//bcax v7.16b,v7.16b,v9.16b,v8.16b
648.inst	0xce3b2508	//bcax v8.16b,v8.16b,v27.16b,   v9.16b
649.inst	0xce266d29	//bcax v9.16b,v9.16b,v6.16b,v27.16b
650
651	eor	v0.16b,v28.16b,v31.16b			// Iota
652
653.inst	0xce2c2d5a	//bcax v26.16b,   v10.16b,v12.16b,v11.16b	// *
654.inst	0xce2d317b	//bcax v27.16b,   v11.16b,v13.16b,v12.16b	// *
655.inst	0xce2e358c	//bcax v12.16b,v12.16b,v14.16b,v13.16b
656.inst	0xce2a39ad	//bcax v13.16b,v13.16b,v10.16b,v14.16b
657.inst	0xce2b29ce	//bcax v14.16b,v14.16b,v11.16b,v10.16b
658
659.inst	0xce3141fc	//bcax v28.16b,   v15.16b,v17.16b,v16.16b	// *
660.inst	0xce32461d	//bcax v29.16b,   v16.16b,v18.16b,v17.16b	// *
661.inst	0xce334a31	//bcax v17.16b,v17.16b,v19.16b,v18.16b
662.inst	0xce2f4e52	//bcax v18.16b,v18.16b,v15.16b,v19.16b
663.inst	0xce303e73	//bcax v19.16b,v19.16b,v16.16b,v15.16b
664
665.inst	0xce36569e	//bcax v30.16b,   v20.16b,v22.16b,v21.16b	// *
666.inst	0xce375abf	//bcax v31.16b,   v21.16b,v23.16b,v22.16b	// *
667.inst	0xce385ed6	//bcax v22.16b,v22.16b,v24.16b,v23.16b
668.inst	0xce3462f7	//bcax v23.16b,v23.16b,v20.16b,v24.16b
669.inst	0xce355318	//bcax v24.16b,v24.16b,v21.16b,v20.16b
670	////////////////////////////////////////////////// Theta
671.inst	0xce056806	//eor3 v6.16b,v0.16b,v5.16b,v26.16b
672.inst	0xce196c2a	//eor3 v10.16b,v1.16b,v25.16b,v27.16b
673.inst	0xce07304b	//eor3 v11.16b,v2.16b,v7.16b,v12.16b
674.inst	0xce08346f	//eor3 v15.16b,v3.16b,v8.16b,v13.16b
675.inst	0xce093890	//eor3 v16.16b,v4.16b,v9.16b,v14.16b
676.inst	0xce1c78c6	//eor3 v6.16b,v6.16b,   v28.16b,v30.16b
677.inst	0xce1d7d4a	//eor3 v10.16b,v10.16b,   v29.16b,v31.16b
678.inst	0xce11596b	//eor3 v11.16b,v11.16b,   v17.16b,v22.16b
679.inst	0xce125def	//eor3 v15.16b,v15.16b,   v18.16b,v23.16b
680.inst	0xce136210	//eor3 v16.16b,v16.16b,   v19.16b,v24.16b
681
682.inst	0xce6b8cd4	//rax1 v20.16b,v6.16b,v11.16b			// D[1]
683.inst	0xce6f8d55	//rax1 v21.16b,v10.16b,v15.16b			// D[2]
684.inst	0xce708d6b	//rax1 v11.16b,v11.16b,v16.16b			// D[3]
685.inst	0xce668def	//rax1 v15.16b,v15.16b,v6.16b			// D[4]
686.inst	0xce6a8e10	//rax1 v16.16b,v16.16b,v10.16b			// D[0]
687
688	////////////////////////////////////////////////// Theta+Rho+Pi
689.inst	0xce945326	//xar v6.16b,   v25.16b,v20.16b,#64-44	// C[0]=A[0][1]
690.inst	0xce8fb139	//xar v25.16b,v9.16b,v15.16b,#64-20
691.inst	0xce950ec9	//xar v9.16b,v22.16b,v21.16b,#64-61
692.inst	0xce8f65d6	//xar v22.16b,v14.16b,v15.16b,#64-39
693.inst	0xce90bbce	//xar v14.16b,v30.16b,v16.16b,#64-18
694
695.inst	0xce95085e	//xar v30.16b,v2.16b,v21.16b,#64-62
696
697.inst	0xce955582	//xar v2.16b,v12.16b,v21.16b,#64-43
698.inst	0xce8b9dac	//xar v12.16b,v13.16b,v11.16b,#64-25
699.inst	0xce8fe26d	//xar v13.16b,v19.16b,v15.16b,#64-8
700.inst	0xce8b22f3	//xar v19.16b,v23.16b,v11.16b,#64-56
701.inst	0xce905f97	//xar v23.16b,v28.16b,v16.16b,#64-41
702
703.inst	0xce8f949c	//xar v28.16b,v4.16b,v15.16b,#64-27
704
705	eor	v0.16b,v0.16b,v16.16b
706	ldr	x11,[x10],#8
707
708.inst	0xce8bae4a	//xar v10.16b,   v18.16b,v11.16b,#64-21	// C[1]=A[0][3]
709.inst	0xce95c632	//xar v18.16b,v17.16b,v21.16b,#64-15
710.inst	0xce94db71	//xar v17.16b,v27.16b,v20.16b,#64-10
711.inst	0xce95e8fb	//xar v27.16b,v7.16b,v21.16b,#64-6
712.inst	0xce90f747	//xar v7.16b,v26.16b,v16.16b,#64-3
713
714.inst	0xce94fc3a	//xar v26.16b,v1.16b,v20.16b,#64-1	// *
715
716.inst	0xce8fcb04	//xar v4.16b,v24.16b,v15.16b,#64-14
717.inst	0xce94fbf8	//xar v24.16b,v31.16b,v20.16b,#64-2
718.inst	0xce8b251f	//xar v31.16b,v8.16b,v11.16b,#64-55
719.inst	0xce944fa8	//xar v8.16b,v29.16b,v20.16b,#64-45
720.inst	0xce9070bd	//xar v29.16b,v5.16b,v16.16b,#64-36
721
722.inst	0xce8b906b	//xar v11.16b,   v3.16b,v11.16b,#64-28	// C[2]=A[1][0]
723
724	////////////////////////////////////////////////// Chi+Iota
725	dup	v21.2d,x11				// borrow C[6]
726.inst	0xce22180f	//bcax v15.16b,   v0.16b,v2.16b,v6.16b	// *
727.inst	0xce2a08c1	//bcax v1.16b,v6.16b,   v10.16b,   v2.16b	// *
728.inst	0xce242842	//bcax v2.16b,v2.16b,v4.16b,v10.16b
729.inst	0xce201143	//bcax v3.16b,v10.16b,   v0.16b,v4.16b
730.inst	0xce260084	//bcax v4.16b,v4.16b,v6.16b,   v0.16b
731
732.inst	0xce276565	//bcax v5.16b,v11.16b,   v7.16b,v25.16b	// *
733.inst	0xce281f26	//bcax v6.16b,   v25.16b,v8.16b,v7.16b	// *
734.inst	0xce2920e7	//bcax v7.16b,v7.16b,v9.16b,v8.16b
735.inst	0xce2b2508	//bcax v8.16b,v8.16b,v11.16b,   v9.16b
736.inst	0xce392d29	//bcax v9.16b,v9.16b,v25.16b,v11.16b
737
738	eor	v0.16b,v15.16b,v21.16b			// Iota
739
740.inst	0xce2c6f4a	//bcax v10.16b,   v26.16b,v12.16b,v27.16b	// *
741.inst	0xce2d336b	//bcax v11.16b,   v27.16b,v13.16b,v12.16b	// *
742.inst	0xce2e358c	//bcax v12.16b,v12.16b,v14.16b,v13.16b
743.inst	0xce3a39ad	//bcax v13.16b,v13.16b,v26.16b,v14.16b
744.inst	0xce3b69ce	//bcax v14.16b,v14.16b,v27.16b,v26.16b
745
746.inst	0xce31778f	//bcax v15.16b,   v28.16b,v17.16b,v29.16b	// *
747.inst	0xce3247b0	//bcax v16.16b,   v29.16b,v18.16b,v17.16b	// *
748.inst	0xce334a31	//bcax v17.16b,v17.16b,v19.16b,v18.16b
749.inst	0xce3c4e52	//bcax v18.16b,v18.16b,v28.16b,v19.16b
750.inst	0xce3d7273	//bcax v19.16b,v19.16b,v29.16b,v28.16b
751
752.inst	0xce367fd4	//bcax v20.16b,   v30.16b,v22.16b,v31.16b	// *
753.inst	0xce375bf5	//bcax v21.16b,   v31.16b,v23.16b,v22.16b	// *
754.inst	0xce385ed6	//bcax v22.16b,v22.16b,v24.16b,v23.16b
755.inst	0xce3e62f7	//bcax v23.16b,v23.16b,v30.16b,v24.16b
756.inst	0xce3f7b18	//bcax v24.16b,v24.16b,v31.16b,v30.16b
757	subs	x9,x9,#1
758	bne	.Loop_ce
759
760	ret
761.size	KeccakF1600_ce,.-KeccakF1600_ce
762
763.type	KeccakF1600_cext,%function
764.align	5
765KeccakF1600_cext:
766.inst	0xd503233f		// paciasp
767	stp	x29,x30,[sp,#-80]!
768	add	x29,sp,#0
769	stp	d8,d9,[sp,#16]		// per ABI requirement
770	stp	d10,d11,[sp,#32]
771	stp	d12,d13,[sp,#48]
772	stp	d14,d15,[sp,#64]
773	ldp	d0,d1,[x0,#8*0]
774	ldp	d2,d3,[x0,#8*2]
775	ldp	d4,d5,[x0,#8*4]
776	ldp	d6,d7,[x0,#8*6]
777	ldp	d8,d9,[x0,#8*8]
778	ldp	d10,d11,[x0,#8*10]
779	ldp	d12,d13,[x0,#8*12]
780	ldp	d14,d15,[x0,#8*14]
781	ldp	d16,d17,[x0,#8*16]
782	ldp	d18,d19,[x0,#8*18]
783	ldp	d20,d21,[x0,#8*20]
784	ldp	d22,d23,[x0,#8*22]
785	ldr	d24,[x0,#8*24]
786	bl	KeccakF1600_ce
787	ldr	x30,[sp,#8]
788	stp	d0,d1,[x0,#8*0]
789	stp	d2,d3,[x0,#8*2]
790	stp	d4,d5,[x0,#8*4]
791	stp	d6,d7,[x0,#8*6]
792	stp	d8,d9,[x0,#8*8]
793	stp	d10,d11,[x0,#8*10]
794	stp	d12,d13,[x0,#8*12]
795	stp	d14,d15,[x0,#8*14]
796	stp	d16,d17,[x0,#8*16]
797	stp	d18,d19,[x0,#8*18]
798	stp	d20,d21,[x0,#8*20]
799	stp	d22,d23,[x0,#8*22]
800	str	d24,[x0,#8*24]
801
802	ldp	d8,d9,[sp,#16]
803	ldp	d10,d11,[sp,#32]
804	ldp	d12,d13,[sp,#48]
805	ldp	d14,d15,[sp,#64]
806	ldr	x29,[sp],#80
807.inst	0xd50323bf		// autiasp
808	ret
809.size	KeccakF1600_cext,.-KeccakF1600_cext
810.globl	SHA3_absorb_cext
811.type	SHA3_absorb_cext,%function
812.align	5
813SHA3_absorb_cext:
814.inst	0xd503233f		// paciasp
815	stp	x29,x30,[sp,#-80]!
816	add	x29,sp,#0
817	stp	d8,d9,[sp,#16]		// per ABI requirement
818	stp	d10,d11,[sp,#32]
819	stp	d12,d13,[sp,#48]
820	stp	d14,d15,[sp,#64]
821	ldp	d0,d1,[x0,#8*0]
822	ldp	d2,d3,[x0,#8*2]
823	ldp	d4,d5,[x0,#8*4]
824	ldp	d6,d7,[x0,#8*6]
825	ldp	d8,d9,[x0,#8*8]
826	ldp	d10,d11,[x0,#8*10]
827	ldp	d12,d13,[x0,#8*12]
828	ldp	d14,d15,[x0,#8*14]
829	ldp	d16,d17,[x0,#8*16]
830	ldp	d18,d19,[x0,#8*18]
831	ldp	d20,d21,[x0,#8*20]
832	ldp	d22,d23,[x0,#8*22]
833	ldr	d24,[x0,#8*24]
834	b	.Loop_absorb_ce
835
836.align	4
837.Loop_absorb_ce:
838	subs	x2,x2,x3		// len - bsz
839	blo	.Labsorbed_ce
840	ldr	d31,[x1],#8		// *inp++
841#ifdef	__AARCH64EB__
842	rev64	v31.16b,v31.16b
843#endif
844	eor	v0.16b,v0.16b,v31.16b
845	cmp	x3,#8*(0+2)
846	blo	.Lprocess_block_ce
847	ldr	d31,[x1],#8		// *inp++
848#ifdef	__AARCH64EB__
849	rev64	v31.16b,v31.16b
850#endif
851	eor	v1.16b,v1.16b,v31.16b
852	beq	.Lprocess_block_ce
853	ldr	d31,[x1],#8		// *inp++
854#ifdef	__AARCH64EB__
855	rev64	v31.16b,v31.16b
856#endif
857	eor	v2.16b,v2.16b,v31.16b
858	cmp	x3,#8*(2+2)
859	blo	.Lprocess_block_ce
860	ldr	d31,[x1],#8		// *inp++
861#ifdef	__AARCH64EB__
862	rev64	v31.16b,v31.16b
863#endif
864	eor	v3.16b,v3.16b,v31.16b
865	beq	.Lprocess_block_ce
866	ldr	d31,[x1],#8		// *inp++
867#ifdef	__AARCH64EB__
868	rev64	v31.16b,v31.16b
869#endif
870	eor	v4.16b,v4.16b,v31.16b
871	cmp	x3,#8*(4+2)
872	blo	.Lprocess_block_ce
873	ldr	d31,[x1],#8		// *inp++
874#ifdef	__AARCH64EB__
875	rev64	v31.16b,v31.16b
876#endif
877	eor	v5.16b,v5.16b,v31.16b
878	beq	.Lprocess_block_ce
879	ldr	d31,[x1],#8		// *inp++
880#ifdef	__AARCH64EB__
881	rev64	v31.16b,v31.16b
882#endif
883	eor	v6.16b,v6.16b,v31.16b
884	cmp	x3,#8*(6+2)
885	blo	.Lprocess_block_ce
886	ldr	d31,[x1],#8		// *inp++
887#ifdef	__AARCH64EB__
888	rev64	v31.16b,v31.16b
889#endif
890	eor	v7.16b,v7.16b,v31.16b
891	beq	.Lprocess_block_ce
892	ldr	d31,[x1],#8		// *inp++
893#ifdef	__AARCH64EB__
894	rev64	v31.16b,v31.16b
895#endif
896	eor	v8.16b,v8.16b,v31.16b
897	cmp	x3,#8*(8+2)
898	blo	.Lprocess_block_ce
899	ldr	d31,[x1],#8		// *inp++
900#ifdef	__AARCH64EB__
901	rev64	v31.16b,v31.16b
902#endif
903	eor	v9.16b,v9.16b,v31.16b
904	beq	.Lprocess_block_ce
905	ldr	d31,[x1],#8		// *inp++
906#ifdef	__AARCH64EB__
907	rev64	v31.16b,v31.16b
908#endif
909	eor	v10.16b,v10.16b,v31.16b
910	cmp	x3,#8*(10+2)
911	blo	.Lprocess_block_ce
912	ldr	d31,[x1],#8		// *inp++
913#ifdef	__AARCH64EB__
914	rev64	v31.16b,v31.16b
915#endif
916	eor	v11.16b,v11.16b,v31.16b
917	beq	.Lprocess_block_ce
918	ldr	d31,[x1],#8		// *inp++
919#ifdef	__AARCH64EB__
920	rev64	v31.16b,v31.16b
921#endif
922	eor	v12.16b,v12.16b,v31.16b
923	cmp	x3,#8*(12+2)
924	blo	.Lprocess_block_ce
925	ldr	d31,[x1],#8		// *inp++
926#ifdef	__AARCH64EB__
927	rev64	v31.16b,v31.16b
928#endif
929	eor	v13.16b,v13.16b,v31.16b
930	beq	.Lprocess_block_ce
931	ldr	d31,[x1],#8		// *inp++
932#ifdef	__AARCH64EB__
933	rev64	v31.16b,v31.16b
934#endif
935	eor	v14.16b,v14.16b,v31.16b
936	cmp	x3,#8*(14+2)
937	blo	.Lprocess_block_ce
938	ldr	d31,[x1],#8		// *inp++
939#ifdef	__AARCH64EB__
940	rev64	v31.16b,v31.16b
941#endif
942	eor	v15.16b,v15.16b,v31.16b
943	beq	.Lprocess_block_ce
944	ldr	d31,[x1],#8		// *inp++
945#ifdef	__AARCH64EB__
946	rev64	v31.16b,v31.16b
947#endif
948	eor	v16.16b,v16.16b,v31.16b
949	cmp	x3,#8*(16+2)
950	blo	.Lprocess_block_ce
951	ldr	d31,[x1],#8		// *inp++
952#ifdef	__AARCH64EB__
953	rev64	v31.16b,v31.16b
954#endif
955	eor	v17.16b,v17.16b,v31.16b
956	beq	.Lprocess_block_ce
957	ldr	d31,[x1],#8		// *inp++
958#ifdef	__AARCH64EB__
959	rev64	v31.16b,v31.16b
960#endif
961	eor	v18.16b,v18.16b,v31.16b
962	cmp	x3,#8*(18+2)
963	blo	.Lprocess_block_ce
964	ldr	d31,[x1],#8		// *inp++
965#ifdef	__AARCH64EB__
966	rev64	v31.16b,v31.16b
967#endif
968	eor	v19.16b,v19.16b,v31.16b
969	beq	.Lprocess_block_ce
970	ldr	d31,[x1],#8		// *inp++
971#ifdef	__AARCH64EB__
972	rev64	v31.16b,v31.16b
973#endif
974	eor	v20.16b,v20.16b,v31.16b
975	cmp	x3,#8*(20+2)
976	blo	.Lprocess_block_ce
977	ldr	d31,[x1],#8		// *inp++
978#ifdef	__AARCH64EB__
979	rev64	v31.16b,v31.16b
980#endif
981	eor	v21.16b,v21.16b,v31.16b
982	beq	.Lprocess_block_ce
983	ldr	d31,[x1],#8		// *inp++
984#ifdef	__AARCH64EB__
985	rev64	v31.16b,v31.16b
986#endif
987	eor	v22.16b,v22.16b,v31.16b
988	cmp	x3,#8*(22+2)
989	blo	.Lprocess_block_ce
990	ldr	d31,[x1],#8		// *inp++
991#ifdef	__AARCH64EB__
992	rev64	v31.16b,v31.16b
993#endif
994	eor	v23.16b,v23.16b,v31.16b
995	beq	.Lprocess_block_ce
996	ldr	d31,[x1],#8		// *inp++
997#ifdef	__AARCH64EB__
998	rev64	v31.16b,v31.16b
999#endif
1000	eor	v24.16b,v24.16b,v31.16b
1001
1002.Lprocess_block_ce:
1003
1004	bl	KeccakF1600_ce
1005
1006	b	.Loop_absorb_ce
1007
1008.align	4
1009.Labsorbed_ce:
1010	stp	d0,d1,[x0,#8*0]
1011	stp	d2,d3,[x0,#8*2]
1012	stp	d4,d5,[x0,#8*4]
1013	stp	d6,d7,[x0,#8*6]
1014	stp	d8,d9,[x0,#8*8]
1015	stp	d10,d11,[x0,#8*10]
1016	stp	d12,d13,[x0,#8*12]
1017	stp	d14,d15,[x0,#8*14]
1018	stp	d16,d17,[x0,#8*16]
1019	stp	d18,d19,[x0,#8*18]
1020	stp	d20,d21,[x0,#8*20]
1021	stp	d22,d23,[x0,#8*22]
1022	str	d24,[x0,#8*24]
1023	add	x0,x2,x3		// return value
1024
1025	ldp	d8,d9,[sp,#16]
1026	ldp	d10,d11,[sp,#32]
1027	ldp	d12,d13,[sp,#48]
1028	ldp	d14,d15,[sp,#64]
1029	ldp	x29,x30,[sp],#80
1030.inst	0xd50323bf		// autiasp
1031	ret
1032.size	SHA3_absorb_cext,.-SHA3_absorb_cext
1033.globl	SHA3_squeeze_cext
1034.type	SHA3_squeeze_cext,%function
1035.align	5
1036SHA3_squeeze_cext:
1037.inst	0xd503233f		// paciasp
1038	stp	x29,x30,[sp,#-16]!
1039	add	x29,sp,#0
1040	mov	x9,x0
1041	mov	x10,x3
1042
1043.Loop_squeeze_ce:
1044	ldr	x4,[x9],#8
1045	cmp	x2,#8
1046	blo	.Lsqueeze_tail_ce
1047#ifdef	__AARCH64EB__
1048	rev	x4,x4
1049#endif
1050	str	x4,[x1],#8
1051	beq	.Lsqueeze_done_ce
1052
1053	sub	x2,x2,#8
1054	subs	x10,x10,#8
1055	bhi	.Loop_squeeze_ce
1056
1057	bl	KeccakF1600_cext
1058	ldr	x30,[sp,#8]
1059	mov	x9,x0
1060	mov	x10,x3
1061	b	.Loop_squeeze_ce
1062
1063.align	4
1064.Lsqueeze_tail_ce:
1065	strb	w4,[x1],#1
1066	lsr	x4,x4,#8
1067	subs	x2,x2,#1
1068	beq	.Lsqueeze_done_ce
1069	strb	w4,[x1],#1
1070	lsr	x4,x4,#8
1071	subs	x2,x2,#1
1072	beq	.Lsqueeze_done_ce
1073	strb	w4,[x1],#1
1074	lsr	x4,x4,#8
1075	subs	x2,x2,#1
1076	beq	.Lsqueeze_done_ce
1077	strb	w4,[x1],#1
1078	lsr	x4,x4,#8
1079	subs	x2,x2,#1
1080	beq	.Lsqueeze_done_ce
1081	strb	w4,[x1],#1
1082	lsr	x4,x4,#8
1083	subs	x2,x2,#1
1084	beq	.Lsqueeze_done_ce
1085	strb	w4,[x1],#1
1086	lsr	x4,x4,#8
1087	subs	x2,x2,#1
1088	beq	.Lsqueeze_done_ce
1089	strb	w4,[x1],#1
1090
1091.Lsqueeze_done_ce:
1092	ldr	x29,[sp],#16
1093.inst	0xd50323bf		// autiasp
1094	ret
1095.size	SHA3_squeeze_cext,.-SHA3_squeeze_cext
1096.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
1097.align	2
1098