• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1// This file is generated from a similarly-named Perl script in the BoringSSL
2// source tree. Do not edit by hand.
3
4#include <ring-core/asm_base.h>
5
6#if !defined(OPENSSL_NO_ASM) && defined(OPENSSL_ARM) && defined(__ELF__)
7.syntax	unified
8
9.arch	armv7-a
10.fpu	neon
11
12#if defined(__thumb2__)
13.thumb
14#else
15.code	32
16#endif
17
18.text
19
20.type	_vpaes_consts,%object
21.align	7	@ totally strategic alignment
22_vpaes_consts:
23.Lk_mc_forward:@ mc_forward
24.quad	0x0407060500030201, 0x0C0F0E0D080B0A09
25.quad	0x080B0A0904070605, 0x000302010C0F0E0D
26.quad	0x0C0F0E0D080B0A09, 0x0407060500030201
27.quad	0x000302010C0F0E0D, 0x080B0A0904070605
28.Lk_mc_backward:@ mc_backward
29.quad	0x0605040702010003, 0x0E0D0C0F0A09080B
30.quad	0x020100030E0D0C0F, 0x0A09080B06050407
31.quad	0x0E0D0C0F0A09080B, 0x0605040702010003
32.quad	0x0A09080B06050407, 0x020100030E0D0C0F
33.Lk_sr:@ sr
34.quad	0x0706050403020100, 0x0F0E0D0C0B0A0908
35.quad	0x030E09040F0A0500, 0x0B06010C07020D08
36.quad	0x0F060D040B020900, 0x070E050C030A0108
37.quad	0x0B0E0104070A0D00, 0x0306090C0F020508
38
39@
40@ "Hot" constants
41@
42.Lk_inv:@ inv, inva
43.quad	0x0E05060F0D080180, 0x040703090A0B0C02
44.quad	0x01040A060F0B0780, 0x030D0E0C02050809
45.Lk_ipt:@ input transform (lo, hi)
46.quad	0xC2B2E8985A2A7000, 0xCABAE09052227808
47.quad	0x4C01307D317C4D00, 0xCD80B1FCB0FDCC81
48.Lk_sbo:@ sbou, sbot
49.quad	0xD0D26D176FBDC700, 0x15AABF7AC502A878
50.quad	0xCFE474A55FBB6A00, 0x8E1E90D1412B35FA
51.Lk_sb1:@ sb1u, sb1t
52.quad	0x3618D415FAE22300, 0x3BF7CCC10D2ED9EF
53.quad	0xB19BE18FCB503E00, 0xA5DF7A6E142AF544
54.Lk_sb2:@ sb2u, sb2t
55.quad	0x69EB88400AE12900, 0xC2A163C8AB82234A
56.quad	0xE27A93C60B712400, 0x5EB7E955BC982FCD
57
58.byte	86,101,99,116,111,114,32,80,101,114,109,117,116,97,116,105,111,110,32,65,69,83,32,102,111,114,32,65,82,77,118,55,32,78,69,79,78,44,32,77,105,107,101,32,72,97,109,98,117,114,103,32,40,83,116,97,110,102,111,114,100,32,85,110,105,118,101,114,115,105,116,121,41,0
59.align	2
60.size	_vpaes_consts,.-_vpaes_consts
61.align	6
62@@
63@@  _aes_preheat
64@@
65@@  Fills q9-q15 as specified below.
66@@
67.type	_vpaes_preheat,%function
68.align	4
69_vpaes_preheat:
70	adr	r10, .Lk_inv
71	vmov.i8	q9, #0x0f		@ .Lk_s0F
72	vld1.64	{q10,q11}, [r10]!	@ .Lk_inv
73	add	r10, r10, #64		@ Skip .Lk_ipt, .Lk_sbo
74	vld1.64	{q12,q13}, [r10]!	@ .Lk_sb1
75	vld1.64	{q14,q15}, [r10]	@ .Lk_sb2
76	bx	lr
77
78@@
79@@  _aes_encrypt_core
80@@
81@@  AES-encrypt q0.
82@@
83@@  Inputs:
84@@     q0 = input
85@@     q9-q15 as in _vpaes_preheat
86@@    [r2] = scheduled keys
87@@
88@@  Output in q0
89@@  Clobbers  q1-q5, r8-r11
90@@  Preserves q6-q8 so you get some local vectors
91@@
92@@
93.type	_vpaes_encrypt_core,%function
94.align	4
95_vpaes_encrypt_core:
96	mov	r9, r2
97	ldr	r8, [r2,#240]		@ pull rounds
98	adr	r11, .Lk_ipt
99	@ vmovdqa	.Lk_ipt(%rip),	%xmm2	# iptlo
100	@ vmovdqa	.Lk_ipt+16(%rip), %xmm3	# ipthi
101	vld1.64	{q2, q3}, [r11]
102	adr	r11, .Lk_mc_forward+16
103	vld1.64	{q5}, [r9]!		@ vmovdqu	(%r9),	%xmm5		# round0 key
104	vand	q1, q0, q9		@ vpand	%xmm9,	%xmm0,	%xmm1
105	vshr.u8	q0, q0, #4		@ vpsrlb	$4,	%xmm0,	%xmm0
106	vtbl.8	d2, {q2}, d2	@ vpshufb	%xmm1,	%xmm2,	%xmm1
107	vtbl.8	d3, {q2}, d3
108	vtbl.8	d4, {q3}, d0	@ vpshufb	%xmm0,	%xmm3,	%xmm2
109	vtbl.8	d5, {q3}, d1
110	veor	q0, q1, q5		@ vpxor	%xmm5,	%xmm1,	%xmm0
111	veor	q0, q0, q2		@ vpxor	%xmm2,	%xmm0,	%xmm0
112
113	@ .Lenc_entry ends with a bnz instruction which is normally paired with
114	@ subs in .Lenc_loop.
115	tst	r8, r8
116	b	.Lenc_entry
117
118.align	4
119.Lenc_loop:
120	@ middle of middle round
121	add	r10, r11, #0x40
122	vtbl.8	d8, {q13}, d4	@ vpshufb	%xmm2,	%xmm13,	%xmm4	# 4 = sb1u
123	vtbl.8	d9, {q13}, d5
124	vld1.64	{q1}, [r11]!		@ vmovdqa	-0x40(%r11,%r10), %xmm1	# .Lk_mc_forward[]
125	vtbl.8	d0, {q12}, d6	@ vpshufb	%xmm3,	%xmm12,	%xmm0	# 0 = sb1t
126	vtbl.8	d1, {q12}, d7
127	veor	q4, q4, q5		@ vpxor		%xmm5,	%xmm4,	%xmm4	# 4 = sb1u + k
128	vtbl.8	d10, {q15}, d4	@ vpshufb	%xmm2,	%xmm15,	%xmm5	# 4 = sb2u
129	vtbl.8	d11, {q15}, d5
130	veor	q0, q0, q4		@ vpxor		%xmm4,	%xmm0,	%xmm0	# 0 = A
131	vtbl.8	d4, {q14}, d6	@ vpshufb	%xmm3,	%xmm14,	%xmm2	# 2 = sb2t
132	vtbl.8	d5, {q14}, d7
133	vld1.64	{q4}, [r10]		@ vmovdqa	(%r11,%r10), %xmm4	# .Lk_mc_backward[]
134	vtbl.8	d6, {q0}, d2	@ vpshufb	%xmm1,	%xmm0,	%xmm3	# 0 = B
135	vtbl.8	d7, {q0}, d3
136	veor	q2, q2, q5		@ vpxor		%xmm5,	%xmm2,	%xmm2	# 2 = 2A
137	@ Write to q5 instead of q0, so the table and destination registers do
138	@ not overlap.
139	vtbl.8	d10, {q0}, d8	@ vpshufb	%xmm4,	%xmm0,	%xmm0	# 3 = D
140	vtbl.8	d11, {q0}, d9
141	veor	q3, q3, q2		@ vpxor		%xmm2,	%xmm3,	%xmm3	# 0 = 2A+B
142	vtbl.8	d8, {q3}, d2	@ vpshufb	%xmm1,	%xmm3,	%xmm4	# 0 = 2B+C
143	vtbl.8	d9, {q3}, d3
144	@ Here we restore the original q0/q5 usage.
145	veor	q0, q5, q3		@ vpxor		%xmm3,	%xmm0,	%xmm0	# 3 = 2A+B+D
146	and	r11, r11, #~(1<<6)	@ and		$0x30,	%r11		# ... mod 4
147	veor	q0, q0, q4		@ vpxor		%xmm4,	%xmm0, %xmm0	# 0 = 2A+3B+C+D
148	subs	r8, r8, #1		@ nr--
149
150.Lenc_entry:
151	@ top of round
152	vand	q1, q0, q9		@ vpand		%xmm0,	%xmm9,	%xmm1   # 0 = k
153	vshr.u8	q0, q0, #4		@ vpsrlb	$4,	%xmm0,	%xmm0	# 1 = i
154	vtbl.8	d10, {q11}, d2	@ vpshufb	%xmm1,	%xmm11,	%xmm5	# 2 = a/k
155	vtbl.8	d11, {q11}, d3
156	veor	q1, q1, q0		@ vpxor		%xmm0,	%xmm1,	%xmm1	# 0 = j
157	vtbl.8	d6, {q10}, d0	@ vpshufb	%xmm0, 	%xmm10,	%xmm3  	# 3 = 1/i
158	vtbl.8	d7, {q10}, d1
159	vtbl.8	d8, {q10}, d2	@ vpshufb	%xmm1, 	%xmm10,	%xmm4  	# 4 = 1/j
160	vtbl.8	d9, {q10}, d3
161	veor	q3, q3, q5		@ vpxor		%xmm5,	%xmm3,	%xmm3	# 3 = iak = 1/i + a/k
162	veor	q4, q4, q5		@ vpxor		%xmm5,	%xmm4,	%xmm4  	# 4 = jak = 1/j + a/k
163	vtbl.8	d4, {q10}, d6	@ vpshufb	%xmm3,	%xmm10,	%xmm2  	# 2 = 1/iak
164	vtbl.8	d5, {q10}, d7
165	vtbl.8	d6, {q10}, d8	@ vpshufb	%xmm4,	%xmm10,	%xmm3	# 3 = 1/jak
166	vtbl.8	d7, {q10}, d9
167	veor	q2, q2, q1		@ vpxor		%xmm1,	%xmm2,	%xmm2  	# 2 = io
168	veor	q3, q3, q0		@ vpxor		%xmm0,	%xmm3,	%xmm3	# 3 = jo
169	vld1.64	{q5}, [r9]!		@ vmovdqu	(%r9),	%xmm5
170	bne	.Lenc_loop
171
172	@ middle of last round
173	add	r10, r11, #0x80
174
175	adr	r11, .Lk_sbo
176	@ Read to q1 instead of q4, so the vtbl.8 instruction below does not
177	@ overlap table and destination registers.
178	vld1.64	{q1}, [r11]!		@ vmovdqa	-0x60(%r10), %xmm4	# 3 : sbou
179	vld1.64	{q0}, [r11]		@ vmovdqa	-0x50(%r10), %xmm0	# 0 : sbot	.Lk_sbo+16
180	vtbl.8	d8, {q1}, d4	@ vpshufb	%xmm2,	%xmm4,	%xmm4	# 4 = sbou
181	vtbl.8	d9, {q1}, d5
182	vld1.64	{q1}, [r10]		@ vmovdqa	0x40(%r11,%r10), %xmm1	# .Lk_sr[]
183	@ Write to q2 instead of q0 below, to avoid overlapping table and
184	@ destination registers.
185	vtbl.8	d4, {q0}, d6	@ vpshufb	%xmm3,	%xmm0,	%xmm0	# 0 = sb1t
186	vtbl.8	d5, {q0}, d7
187	veor	q4, q4, q5		@ vpxor	%xmm5,	%xmm4,	%xmm4	# 4 = sb1u + k
188	veor	q2, q2, q4		@ vpxor	%xmm4,	%xmm0,	%xmm0	# 0 = A
189	@ Here we restore the original q0/q2 usage.
190	vtbl.8	d0, {q2}, d2	@ vpshufb	%xmm1,	%xmm0,	%xmm0
191	vtbl.8	d1, {q2}, d3
192	bx	lr
193.size	_vpaes_encrypt_core,.-_vpaes_encrypt_core
194
195.globl	vpaes_encrypt
196.hidden	vpaes_encrypt
197.type	vpaes_encrypt,%function
198.align	4
199vpaes_encrypt:
200	@ _vpaes_encrypt_core uses r8-r11. Round up to r7-r11 to maintain stack
201	@ alignment.
202	stmdb	sp!, {r7,r8,r9,r10,r11,lr}
203	@ _vpaes_encrypt_core uses q4-q5 (d8-d11), which are callee-saved.
204	vstmdb	sp!, {d8,d9,d10,d11}
205
206	vld1.64	{q0}, [r0]
207	bl	_vpaes_preheat
208	bl	_vpaes_encrypt_core
209	vst1.64	{q0}, [r1]
210
211	vldmia	sp!, {d8,d9,d10,d11}
212	ldmia	sp!, {r7,r8,r9,r10,r11, pc}	@ return
213.size	vpaes_encrypt,.-vpaes_encrypt
214@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
215@@                                                    @@
216@@                  AES key schedule                  @@
217@@                                                    @@
218@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
219
220@ This function diverges from both x86_64 and armv7 in which constants are
221@ pinned. x86_64 has a common preheat function for all operations. aarch64
222@ separates them because it has enough registers to pin nearly all constants.
223@ armv7 does not have enough registers, but needing explicit loads and stores
224@ also complicates using x86_64's register allocation directly.
225@
226@ We pin some constants for convenience and leave q14 and q15 free to load
227@ others on demand.
228
229@
230@  Key schedule constants
231@
232.type	_vpaes_key_consts,%object
233.align	4
234_vpaes_key_consts:
235.Lk_rcon:@ rcon
236.quad	0x1F8391B9AF9DEEB6, 0x702A98084D7C7D81
237
238.Lk_opt:@ output transform
239.quad	0xFF9F4929D6B66000, 0xF7974121DEBE6808
240.quad	0x01EDBD5150BCEC00, 0xE10D5DB1B05C0CE0
241.Lk_deskew:@ deskew tables: inverts the sbox's "skew"
242.quad	0x07E4A34047A4E300, 0x1DFEB95A5DBEF91A
243.quad	0x5F36B5DC83EA6900, 0x2841C2ABF49D1E77
244.size	_vpaes_key_consts,.-_vpaes_key_consts
245
246.type	_vpaes_key_preheat,%function
247.align	4
248_vpaes_key_preheat:
249	adr	r11, .Lk_rcon
250	vmov.i8	q12, #0x5b			@ .Lk_s63
251	adr	r10, .Lk_inv			@ Must be aligned to 8 mod 16.
252	vmov.i8	q9, #0x0f			@ .Lk_s0F
253	vld1.64	{q10,q11}, [r10]		@ .Lk_inv
254	vld1.64	{q8}, [r11]			@ .Lk_rcon
255	bx	lr
256.size	_vpaes_key_preheat,.-_vpaes_key_preheat
257
258.type	_vpaes_schedule_core,%function
259.align	4
260_vpaes_schedule_core:
261	@ We only need to save lr, but ARM requires an 8-byte stack alignment,
262	@ so save an extra register.
263	stmdb	sp!, {r3,lr}
264
265	bl	_vpaes_key_preheat	@ load the tables
266
267	adr	r11, .Lk_ipt		@ Must be aligned to 8 mod 16.
268	vld1.64	{q0}, [r0]!		@ vmovdqu	(%rdi),	%xmm0		# load key (unaligned)
269
270	@ input transform
271	@ Use q4 here rather than q3 so .Lschedule_am_decrypting does not
272	@ overlap table and destination.
273	vmov	q4, q0			@ vmovdqa	%xmm0,	%xmm3
274	bl	_vpaes_schedule_transform
275	adr	r10, .Lk_sr		@ Must be aligned to 8 mod 16.
276	vmov	q7, q0			@ vmovdqa	%xmm0,	%xmm7
277
278	add	r8, r8, r10
279
280	@ encrypting, output zeroth round key after transform
281	vst1.64	{q0}, [r2]		@ vmovdqu	%xmm0,	(%rdx)
282
283	@ *ring*: Decryption removed.
284
285.Lschedule_go:
286	cmp	r1, #192		@ cmp	$192,	%esi
287	bhi	.Lschedule_256
288	@ 128: fall though
289
290@@
291@@  .schedule_128
292@@
293@@  128-bit specific part of key schedule.
294@@
295@@  This schedule is really simple, because all its parts
296@@  are accomplished by the subroutines.
297@@
298.Lschedule_128:
299	mov	r0, #10		@ mov	$10, %esi
300
301.Loop_schedule_128:
302	bl	_vpaes_schedule_round
303	subs	r0, r0, #1		@ dec	%esi
304	beq	.Lschedule_mangle_last
305	bl	_vpaes_schedule_mangle	@ write output
306	b	.Loop_schedule_128
307
308@@
309@@  .aes_schedule_256
310@@
311@@  256-bit specific part of key schedule.
312@@
313@@  The structure here is very similar to the 128-bit
314@@  schedule, but with an additional "low side" in
315@@  q6.  The low side's rounds are the same as the
316@@  high side's, except no rcon and no rotation.
317@@
318.align	4
319.Lschedule_256:
320	vld1.64	{q0}, [r0]			@ vmovdqu	16(%rdi),%xmm0		# load key part 2 (unaligned)
321	bl	_vpaes_schedule_transform	@ input transform
322	mov	r0, #7			@ mov	$7, %esi
323
324.Loop_schedule_256:
325	bl	_vpaes_schedule_mangle		@ output low result
326	vmov	q6, q0				@ vmovdqa	%xmm0,	%xmm6		# save cur_lo in xmm6
327
328	@ high round
329	bl	_vpaes_schedule_round
330	subs	r0, r0, #1			@ dec	%esi
331	beq	.Lschedule_mangle_last
332	bl	_vpaes_schedule_mangle
333
334	@ low round. swap xmm7 and xmm6
335	vdup.32	q0, d1[1]		@ vpshufd	$0xFF,	%xmm0,	%xmm0
336	vmov.i8	q4, #0
337	vmov	q5, q7			@ vmovdqa	%xmm7,	%xmm5
338	vmov	q7, q6			@ vmovdqa	%xmm6,	%xmm7
339	bl	_vpaes_schedule_low_round
340	vmov	q7, q5			@ vmovdqa	%xmm5,	%xmm7
341
342	b	.Loop_schedule_256
343
344@@
345@@  .aes_schedule_mangle_last
346@@
347@@  Mangler for last round of key schedule
348@@  Mangles q0
349@@    when encrypting, outputs out(q0) ^ 63
350@@    when decrypting, outputs unskew(q0)
351@@
352@@  Always called right before return... jumps to cleanup and exits
353@@
354.align	4
355.Lschedule_mangle_last:
356	@ schedule last round key from xmm0
357	adr	r11, .Lk_deskew			@ lea	.Lk_deskew(%rip),%r11	# prepare to deskew
358
359	@ encrypting
360	vld1.64	{q1}, [r8]		@ vmovdqa	(%r8,%r10),%xmm1
361	adr	r11, .Lk_opt		@ lea		.Lk_opt(%rip),	%r11		# prepare to output transform
362	add	r2, r2, #32		@ add		$32,	%rdx
363	vmov	q2, q0
364	vtbl.8	d0, {q2}, d2	@ vpshufb	%xmm1,	%xmm0,	%xmm0		# output permute
365	vtbl.8	d1, {q2}, d3
366
367.Lschedule_mangle_last_dec:
368	sub	r2, r2, #16			@ add	$-16,	%rdx
369	veor	q0, q0, q12			@ vpxor	.Lk_s63(%rip),	%xmm0,	%xmm0
370	bl	_vpaes_schedule_transform	@ output transform
371	vst1.64	{q0}, [r2]			@ vmovdqu	%xmm0,	(%rdx)		# save last key
372
373	@ cleanup
374	veor	q0, q0, q0		@ vpxor	%xmm0,	%xmm0,	%xmm0
375	veor	q1, q1, q1		@ vpxor	%xmm1,	%xmm1,	%xmm1
376	veor	q2, q2, q2		@ vpxor	%xmm2,	%xmm2,	%xmm2
377	veor	q3, q3, q3		@ vpxor	%xmm3,	%xmm3,	%xmm3
378	veor	q4, q4, q4		@ vpxor	%xmm4,	%xmm4,	%xmm4
379	veor	q5, q5, q5		@ vpxor	%xmm5,	%xmm5,	%xmm5
380	veor	q6, q6, q6		@ vpxor	%xmm6,	%xmm6,	%xmm6
381	veor	q7, q7, q7		@ vpxor	%xmm7,	%xmm7,	%xmm7
382	ldmia	sp!, {r3,pc}		@ return
383.size	_vpaes_schedule_core,.-_vpaes_schedule_core
384
385@@
386@@  .aes_schedule_round
387@@
388@@  Runs one main round of the key schedule on q0, q7
389@@
390@@  Specifically, runs subbytes on the high dword of q0
391@@  then rotates it by one byte and xors into the low dword of
392@@  q7.
393@@
394@@  Adds rcon from low byte of q8, then rotates q8 for
395@@  next rcon.
396@@
397@@  Smears the dwords of q7 by xoring the low into the
398@@  second low, result into third, result into highest.
399@@
400@@  Returns results in q7 = q0.
401@@  Clobbers q1-q4, r11.
402@@
403.type	_vpaes_schedule_round,%function
404.align	4
405_vpaes_schedule_round:
406	@ extract rcon from xmm8
407	vmov.i8	q4, #0				@ vpxor		%xmm4,	%xmm4,	%xmm4
408	vext.8	q1, q8, q4, #15		@ vpalignr	$15,	%xmm8,	%xmm4,	%xmm1
409	vext.8	q8, q8, q8, #15	@ vpalignr	$15,	%xmm8,	%xmm8,	%xmm8
410	veor	q7, q7, q1			@ vpxor		%xmm1,	%xmm7,	%xmm7
411
412	@ rotate
413	vdup.32	q0, d1[1]			@ vpshufd	$0xFF,	%xmm0,	%xmm0
414	vext.8	q0, q0, q0, #1			@ vpalignr	$1,	%xmm0,	%xmm0,	%xmm0
415
416	@ fall through...
417
418	@ low round: same as high round, but no rotation and no rcon.
419_vpaes_schedule_low_round:
420	@ The x86_64 version pins .Lk_sb1 in %xmm13 and .Lk_sb1+16 in %xmm12.
421	@ We pin other values in _vpaes_key_preheat, so load them now.
422	adr	r11, .Lk_sb1
423	vld1.64	{q14,q15}, [r11]
424
425	@ smear xmm7
426	vext.8	q1, q4, q7, #12			@ vpslldq	$4,	%xmm7,	%xmm1
427	veor	q7, q7, q1			@ vpxor	%xmm1,	%xmm7,	%xmm7
428	vext.8	q4, q4, q7, #8			@ vpslldq	$8,	%xmm7,	%xmm4
429
430	@ subbytes
431	vand	q1, q0, q9			@ vpand		%xmm9,	%xmm0,	%xmm1		# 0 = k
432	vshr.u8	q0, q0, #4			@ vpsrlb	$4,	%xmm0,	%xmm0		# 1 = i
433	veor	q7, q7, q4			@ vpxor		%xmm4,	%xmm7,	%xmm7
434	vtbl.8	d4, {q11}, d2		@ vpshufb	%xmm1,	%xmm11,	%xmm2		# 2 = a/k
435	vtbl.8	d5, {q11}, d3
436	veor	q1, q1, q0			@ vpxor		%xmm0,	%xmm1,	%xmm1		# 0 = j
437	vtbl.8	d6, {q10}, d0		@ vpshufb	%xmm0, 	%xmm10,	%xmm3		# 3 = 1/i
438	vtbl.8	d7, {q10}, d1
439	veor	q3, q3, q2			@ vpxor		%xmm2,	%xmm3,	%xmm3		# 3 = iak = 1/i + a/k
440	vtbl.8	d8, {q10}, d2		@ vpshufb	%xmm1,	%xmm10,	%xmm4		# 4 = 1/j
441	vtbl.8	d9, {q10}, d3
442	veor	q7, q7, q12			@ vpxor		.Lk_s63(%rip),	%xmm7,	%xmm7
443	vtbl.8	d6, {q10}, d6		@ vpshufb	%xmm3,	%xmm10,	%xmm3		# 2 = 1/iak
444	vtbl.8	d7, {q10}, d7
445	veor	q4, q4, q2			@ vpxor		%xmm2,	%xmm4,	%xmm4		# 4 = jak = 1/j + a/k
446	vtbl.8	d4, {q10}, d8		@ vpshufb	%xmm4,	%xmm10,	%xmm2		# 3 = 1/jak
447	vtbl.8	d5, {q10}, d9
448	veor	q3, q3, q1			@ vpxor		%xmm1,	%xmm3,	%xmm3		# 2 = io
449	veor	q2, q2, q0			@ vpxor		%xmm0,	%xmm2,	%xmm2		# 3 = jo
450	vtbl.8	d8, {q15}, d6		@ vpshufb	%xmm3,	%xmm13,	%xmm4		# 4 = sbou
451	vtbl.8	d9, {q15}, d7
452	vtbl.8	d2, {q14}, d4		@ vpshufb	%xmm2,	%xmm12,	%xmm1		# 0 = sb1t
453	vtbl.8	d3, {q14}, d5
454	veor	q1, q1, q4			@ vpxor		%xmm4,	%xmm1,	%xmm1		# 0 = sbox output
455
456	@ add in smeared stuff
457	veor	q0, q1, q7			@ vpxor	%xmm7,	%xmm1,	%xmm0
458	veor	q7, q1, q7			@ vmovdqa	%xmm0,	%xmm7
459	bx	lr
460.size	_vpaes_schedule_round,.-_vpaes_schedule_round
461
462@@
463@@  .aes_schedule_transform
464@@
465@@  Linear-transform q0 according to tables at [r11]
466@@
467@@  Requires that q9 = 0x0F0F... as in preheat
468@@  Output in q0
469@@  Clobbers q1, q2, q14, q15
470@@
471.type	_vpaes_schedule_transform,%function
472.align	4
473_vpaes_schedule_transform:
474	vld1.64	{q14,q15}, [r11]	@ vmovdqa	(%r11),	%xmm2 	# lo
475					@ vmovdqa	16(%r11),	%xmm1 # hi
476	vand	q1, q0, q9		@ vpand	%xmm9,	%xmm0,	%xmm1
477	vshr.u8	q0, q0, #4		@ vpsrlb	$4,	%xmm0,	%xmm0
478	vtbl.8	d4, {q14}, d2	@ vpshufb	%xmm1,	%xmm2,	%xmm2
479	vtbl.8	d5, {q14}, d3
480	vtbl.8	d0, {q15}, d0	@ vpshufb	%xmm0,	%xmm1,	%xmm0
481	vtbl.8	d1, {q15}, d1
482	veor	q0, q0, q2		@ vpxor	%xmm2,	%xmm0,	%xmm0
483	bx	lr
484.size	_vpaes_schedule_transform,.-_vpaes_schedule_transform
485
486@@
487@@  .aes_schedule_mangle
488@@
489@@  Mangles q0 from (basis-transformed) standard version
490@@  to our version.
491@@
492@@  On encrypt,
493@@    xor with 0x63
494@@    multiply by circulant 0,1,1,1
495@@    apply shiftrows transform
496@@
497@@  On decrypt,
498@@    xor with 0x63
499@@    multiply by "inverse mixcolumns" circulant E,B,D,9
500@@    deskew
501@@    apply shiftrows transform
502@@
503@@
504@@  Writes out to [r2], and increments or decrements it
505@@  Keeps track of round number mod 4 in r8
506@@  Preserves q0
507@@  Clobbers q1-q5
508@@
509.type	_vpaes_schedule_mangle,%function
510.align	4
511_vpaes_schedule_mangle:
512	tst	r3, r3
513	vmov	q4, q0			@ vmovdqa	%xmm0,	%xmm4	# save xmm0 for later
514	adr	r11, .Lk_mc_forward	@ Must be aligned to 8 mod 16.
515	vld1.64	{q5}, [r11]		@ vmovdqa	.Lk_mc_forward(%rip),%xmm5
516
517	@ encrypting
518	@ Write to q2 so we do not overlap table and destination below.
519	veor	q2, q0, q12		@ vpxor		.Lk_s63(%rip),	%xmm0,	%xmm4
520	add	r2, r2, #16		@ add		$16,	%rdx
521	vtbl.8	d8, {q2}, d10	@ vpshufb	%xmm5,	%xmm4,	%xmm4
522	vtbl.8	d9, {q2}, d11
523	vtbl.8	d2, {q4}, d10	@ vpshufb	%xmm5,	%xmm4,	%xmm1
524	vtbl.8	d3, {q4}, d11
525	vtbl.8	d6, {q1}, d10	@ vpshufb	%xmm5,	%xmm1,	%xmm3
526	vtbl.8	d7, {q1}, d11
527	veor	q4, q4, q1		@ vpxor		%xmm1,	%xmm4,	%xmm4
528	vld1.64	{q1}, [r8]		@ vmovdqa	(%r8,%r10),	%xmm1
529	veor	q3, q3, q4		@ vpxor		%xmm4,	%xmm3,	%xmm3
530
531.Lschedule_mangle_both:
532	@ Write to q2 so table and destination do not overlap.
533	vtbl.8	d4, {q3}, d2	@ vpshufb	%xmm1,	%xmm3,	%xmm3
534	vtbl.8	d5, {q3}, d3
535	add	r8, r8, #64-16		@ add	$-16,	%r8
536	and	r8, r8, #~(1<<6)	@ and	$0x30,	%r8
537	vst1.64	{q2}, [r2]		@ vmovdqu	%xmm3,	(%rdx)
538	bx	lr
539.size	_vpaes_schedule_mangle,.-_vpaes_schedule_mangle
540
541.globl	vpaes_set_encrypt_key
542.hidden	vpaes_set_encrypt_key
543.type	vpaes_set_encrypt_key,%function
544.align	4
545vpaes_set_encrypt_key:
546	stmdb	sp!, {r7,r8,r9,r10,r11, lr}
547	vstmdb	sp!, {d8,d9,d10,d11,d12,d13,d14,d15}
548
549	lsr	r9, r1, #5		@ shr	$5,%eax
550	add	r9, r9, #5		@ $5,%eax
551	str	r9, [r2,#240]		@ mov	%eax,240(%rdx)	# AES_KEY->rounds = nbits/32+5;
552
553	mov	r3, #0		@ mov	$0,%ecx
554	mov	r8, #0x30		@ mov	$0x30,%r8d
555	bl	_vpaes_schedule_core
556	eor	r0, r0, r0
557
558	vldmia	sp!, {d8,d9,d10,d11,d12,d13,d14,d15}
559	ldmia	sp!, {r7,r8,r9,r10,r11, pc}	@ return
560.size	vpaes_set_encrypt_key,.-vpaes_set_encrypt_key
561
562@ Additional constants for converting to bsaes.
563.type	_vpaes_convert_consts,%object
564.align	4
565_vpaes_convert_consts:
566@ .Lk_opt_then_skew applies skew(opt(x)) XOR 0x63, where skew is the linear
567@ transform in the AES S-box. 0x63 is incorporated into the low half of the
568@ table. This was computed with the following script:
569@
570@   def u64s_to_u128(x, y):
571@       return x | (y << 64)
572@   def u128_to_u64s(w):
573@       return w & ((1<<64)-1), w >> 64
574@   def get_byte(w, i):
575@       return (w >> (i*8)) & 0xff
576@   def apply_table(table, b):
577@       lo = b & 0xf
578@       hi = b >> 4
579@       return get_byte(table[0], lo) ^ get_byte(table[1], hi)
580@   def opt(b):
581@       table = [
582@           u64s_to_u128(0xFF9F4929D6B66000, 0xF7974121DEBE6808),
583@           u64s_to_u128(0x01EDBD5150BCEC00, 0xE10D5DB1B05C0CE0),
584@       ]
585@       return apply_table(table, b)
586@   def rot_byte(b, n):
587@       return 0xff & ((b << n) | (b >> (8-n)))
588@   def skew(x):
589@       return (x ^ rot_byte(x, 1) ^ rot_byte(x, 2) ^ rot_byte(x, 3) ^
590@               rot_byte(x, 4))
591@   table = [0, 0]
592@   for i in range(16):
593@       table[0] |= (skew(opt(i)) ^ 0x63) << (i*8)
594@       table[1] |= skew(opt(i<<4)) << (i*8)
595@   print("	.quad	0x%016x, 0x%016x" % u128_to_u64s(table[0]))
596@   print("	.quad	0x%016x, 0x%016x" % u128_to_u64s(table[1]))
597.Lk_opt_then_skew:
598.quad	0x9cb8436798bc4763, 0x6440bb9f6044bf9b
599.quad	0x1f30062936192f00, 0xb49bad829db284ab
600
601@ void vpaes_encrypt_key_to_bsaes(AES_KEY *bsaes, const AES_KEY *vpaes);
602.globl	vpaes_encrypt_key_to_bsaes
603.hidden	vpaes_encrypt_key_to_bsaes
604.type	vpaes_encrypt_key_to_bsaes,%function
605.align	4
606vpaes_encrypt_key_to_bsaes:
607	stmdb	sp!, {r11, lr}
608
609	@ See _vpaes_schedule_core for the key schedule logic. In particular,
610	@ _vpaes_schedule_transform(.Lk_ipt) (section 2.2 of the paper),
611	@ _vpaes_schedule_mangle (section 4.3), and .Lschedule_mangle_last
612	@ contain the transformations not in the bsaes representation. This
613	@ function inverts those transforms.
614	@
615	@ Note also that bsaes-armv7.pl expects aes-armv4.pl's key
616	@ representation, which does not match the other aes_nohw_*
617	@ implementations. The ARM aes_nohw_* stores each 32-bit word
618	@ byteswapped, as a convenience for (unsupported) big-endian ARM, at the
619	@ cost of extra REV and VREV32 operations in little-endian ARM.
620
621	vmov.i8	q9, #0x0f		@ Required by _vpaes_schedule_transform
622	adr	r2, .Lk_mc_forward	@ Must be aligned to 8 mod 16.
623	add	r3, r2, 0x90		@ .Lk_sr+0x10-.Lk_mc_forward = 0x90 (Apple's toolchain doesn't support the expression)
624
625	vld1.64	{q12}, [r2]
626	vmov.i8	q10, #0x5b		@ .Lk_s63 from vpaes-x86_64
627	adr	r11, .Lk_opt		@ Must be aligned to 8 mod 16.
628	vmov.i8	q11, #0x63		@ .LK_s63 without .Lk_ipt applied
629
630	@ vpaes stores one fewer round count than bsaes, but the number of keys
631	@ is the same.
632	ldr	r2, [r1,#240]
633	add	r2, r2, #1
634	str	r2, [r0,#240]
635
636	@ The first key is transformed with _vpaes_schedule_transform(.Lk_ipt).
637	@ Invert this with .Lk_opt.
638	vld1.64	{q0}, [r1]!
639	bl	_vpaes_schedule_transform
640	vrev32.8	q0, q0
641	vst1.64	{q0}, [r0]!
642
643	@ The middle keys have _vpaes_schedule_transform(.Lk_ipt) applied,
644	@ followed by _vpaes_schedule_mangle. _vpaes_schedule_mangle XORs 0x63,
645	@ multiplies by the circulant 0,1,1,1, then applies ShiftRows.
646.Loop_enc_key_to_bsaes:
647	vld1.64	{q0}, [r1]!
648
649	@ Invert the ShiftRows step (see .Lschedule_mangle_both). Note we cycle
650	@ r3 in the opposite direction and start at .Lk_sr+0x10 instead of 0x30.
651	@ We use r3 rather than r8 to avoid a callee-saved register.
652	vld1.64	{q1}, [r3]
653	vtbl.8	d4, {q0}, d2
654	vtbl.8	d5, {q0}, d3
655	add	r3, r3, #16
656	and	r3, r3, #~(1<<6)
657	vmov	q0, q2
658
659	@ Handle the last key differently.
660	subs	r2, r2, #1
661	beq	.Loop_enc_key_to_bsaes_last
662
663	@ Multiply by the circulant. This is its own inverse.
664	vtbl.8	d2, {q0}, d24
665	vtbl.8	d3, {q0}, d25
666	vmov	q0, q1
667	vtbl.8	d4, {q1}, d24
668	vtbl.8	d5, {q1}, d25
669	veor	q0, q0, q2
670	vtbl.8	d2, {q2}, d24
671	vtbl.8	d3, {q2}, d25
672	veor	q0, q0, q1
673
674	@ XOR and finish.
675	veor	q0, q0, q10
676	bl	_vpaes_schedule_transform
677	vrev32.8	q0, q0
678	vst1.64	{q0}, [r0]!
679	b	.Loop_enc_key_to_bsaes
680
681.Loop_enc_key_to_bsaes_last:
682	@ The final key does not have a basis transform (note
683	@ .Lschedule_mangle_last inverts the original transform). It only XORs
684	@ 0x63 and applies ShiftRows. The latter was already inverted in the
685	@ loop. Note that, because we act on the original representation, we use
686	@ q11, not q10.
687	veor	q0, q0, q11
688	vrev32.8	q0, q0
689	vst1.64	{q0}, [r0]
690
691	@ Wipe registers which contained key material.
692	veor	q0, q0, q0
693	veor	q1, q1, q1
694	veor	q2, q2, q2
695
696	ldmia	sp!, {r11, pc}	@ return
697.size	vpaes_encrypt_key_to_bsaes,.-vpaes_encrypt_key_to_bsaes
698.globl	vpaes_ctr32_encrypt_blocks
699.hidden	vpaes_ctr32_encrypt_blocks
700.type	vpaes_ctr32_encrypt_blocks,%function
701.align	4
702vpaes_ctr32_encrypt_blocks:
703	mov	ip, sp
704	stmdb	sp!, {r7,r8,r9,r10,r11, lr}
705	@ This function uses q4-q7 (d8-d15), which are callee-saved.
706	vstmdb	sp!, {d8,d9,d10,d11,d12,d13,d14,d15}
707
708	cmp	r2, #0
709	@ r8 is passed on the stack.
710	ldr	r8, [ip]
711	beq	.Lctr32_done
712
713	@ _vpaes_encrypt_core expects the key in r2, so swap r2 and r3.
714	mov	r9, r3
715	mov	r3, r2
716	mov	r2, r9
717
718	@ Load the IV and counter portion.
719	ldr	r7, [r8, #12]
720	vld1.8	{q7}, [r8]
721
722	bl	_vpaes_preheat
723	rev	r7, r7		@ The counter is big-endian.
724
725.Lctr32_loop:
726	vmov	q0, q7
727	vld1.8	{q6}, [r0]!		@ .Load input ahead of time
728	bl	_vpaes_encrypt_core
729	veor	q0, q0, q6		@ XOR input and result
730	vst1.8	{q0}, [r1]!
731	subs	r3, r3, #1
732	@ Update the counter.
733	add	r7, r7, #1
734	rev	r9, r7
735	vmov.32	d15[1], r9
736	bne	.Lctr32_loop
737
738.Lctr32_done:
739	vldmia	sp!, {d8,d9,d10,d11,d12,d13,d14,d15}
740	ldmia	sp!, {r7,r8,r9,r10,r11, pc}	@ return
741.size	vpaes_ctr32_encrypt_blocks,.-vpaes_ctr32_encrypt_blocks
742#endif  // !OPENSSL_NO_ASM && defined(OPENSSL_ARM) && defined(__ELF__)
743