• 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#if !defined(__has_feature)
5#define __has_feature(x) 0
6#endif
7#if __has_feature(memory_sanitizer) && !defined(OPENSSL_NO_ASM)
8#define OPENSSL_NO_ASM
9#endif
10
11#if !defined(OPENSSL_NO_ASM)
12#if defined(__arm__)
13#include "ring_core_generated/prefix_symbols_asm.h"
14.syntax	unified
15
16.arch	armv7-a
17.fpu	neon
18
19#if defined(__thumb2__)
20.thumb
21#else
22.code	32
23#endif
24
25.text
26
27.type	_vpaes_consts,%object
28.align	7	@ totally strategic alignment
29_vpaes_consts:
30.Lk_mc_forward:@ mc_forward
31.quad	0x0407060500030201, 0x0C0F0E0D080B0A09
32.quad	0x080B0A0904070605, 0x000302010C0F0E0D
33.quad	0x0C0F0E0D080B0A09, 0x0407060500030201
34.quad	0x000302010C0F0E0D, 0x080B0A0904070605
35.Lk_mc_backward:@ mc_backward
36.quad	0x0605040702010003, 0x0E0D0C0F0A09080B
37.quad	0x020100030E0D0C0F, 0x0A09080B06050407
38.quad	0x0E0D0C0F0A09080B, 0x0605040702010003
39.quad	0x0A09080B06050407, 0x020100030E0D0C0F
40.Lk_sr:@ sr
41.quad	0x0706050403020100, 0x0F0E0D0C0B0A0908
42.quad	0x030E09040F0A0500, 0x0B06010C07020D08
43.quad	0x0F060D040B020900, 0x070E050C030A0108
44.quad	0x0B0E0104070A0D00, 0x0306090C0F020508
45
46@
47@ "Hot" constants
48@
49.Lk_inv:@ inv, inva
50.quad	0x0E05060F0D080180, 0x040703090A0B0C02
51.quad	0x01040A060F0B0780, 0x030D0E0C02050809
52.Lk_ipt:@ input transform (lo, hi)
53.quad	0xC2B2E8985A2A7000, 0xCABAE09052227808
54.quad	0x4C01307D317C4D00, 0xCD80B1FCB0FDCC81
55.Lk_sbo:@ sbou, sbot
56.quad	0xD0D26D176FBDC700, 0x15AABF7AC502A878
57.quad	0xCFE474A55FBB6A00, 0x8E1E90D1412B35FA
58.Lk_sb1:@ sb1u, sb1t
59.quad	0x3618D415FAE22300, 0x3BF7CCC10D2ED9EF
60.quad	0xB19BE18FCB503E00, 0xA5DF7A6E142AF544
61.Lk_sb2:@ sb2u, sb2t
62.quad	0x69EB88400AE12900, 0xC2A163C8AB82234A
63.quad	0xE27A93C60B712400, 0x5EB7E955BC982FCD
64
65.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
66.align	2
67.size	_vpaes_consts,.-_vpaes_consts
68.align	6
69@@
70@@  _aes_preheat
71@@
72@@  Fills q9-q15 as specified below.
73@@
74.type	_vpaes_preheat,%function
75.align	4
76_vpaes_preheat:
77	adr	r10, .Lk_inv
78	vmov.i8	q9, #0x0f		@ .Lk_s0F
79	vld1.64	{q10,q11}, [r10]!	@ .Lk_inv
80	add	r10, r10, #64		@ Skip .Lk_ipt, .Lk_sbo
81	vld1.64	{q12,q13}, [r10]!	@ .Lk_sb1
82	vld1.64	{q14,q15}, [r10]	@ .Lk_sb2
83	bx	lr
84
85@@
86@@  _aes_encrypt_core
87@@
88@@  AES-encrypt q0.
89@@
90@@  Inputs:
91@@     q0 = input
92@@     q9-q15 as in _vpaes_preheat
93@@    [r2] = scheduled keys
94@@
95@@  Output in q0
96@@  Clobbers  q1-q5, r8-r11
97@@  Preserves q6-q8 so you get some local vectors
98@@
99@@
100.type	_vpaes_encrypt_core,%function
101.align	4
102_vpaes_encrypt_core:
103	mov	r9, r2
104	ldr	r8, [r2,#240]		@ pull rounds
105	adr	r11, .Lk_ipt
106	@ vmovdqa	.Lk_ipt(%rip),	%xmm2	# iptlo
107	@ vmovdqa	.Lk_ipt+16(%rip), %xmm3	# ipthi
108	vld1.64	{q2, q3}, [r11]
109	adr	r11, .Lk_mc_forward+16
110	vld1.64	{q5}, [r9]!		@ vmovdqu	(%r9),	%xmm5		# round0 key
111	vand	q1, q0, q9		@ vpand	%xmm9,	%xmm0,	%xmm1
112	vshr.u8	q0, q0, #4		@ vpsrlb	$4,	%xmm0,	%xmm0
113	vtbl.8	d2, {q2}, d2	@ vpshufb	%xmm1,	%xmm2,	%xmm1
114	vtbl.8	d3, {q2}, d3
115	vtbl.8	d4, {q3}, d0	@ vpshufb	%xmm0,	%xmm3,	%xmm2
116	vtbl.8	d5, {q3}, d1
117	veor	q0, q1, q5		@ vpxor	%xmm5,	%xmm1,	%xmm0
118	veor	q0, q0, q2		@ vpxor	%xmm2,	%xmm0,	%xmm0
119
120	@ .Lenc_entry ends with a bnz instruction which is normally paired with
121	@ subs in .Lenc_loop.
122	tst	r8, r8
123	b	.Lenc_entry
124
125.align	4
126.Lenc_loop:
127	@ middle of middle round
128	add	r10, r11, #0x40
129	vtbl.8	d8, {q13}, d4	@ vpshufb	%xmm2,	%xmm13,	%xmm4	# 4 = sb1u
130	vtbl.8	d9, {q13}, d5
131	vld1.64	{q1}, [r11]!		@ vmovdqa	-0x40(%r11,%r10), %xmm1	# .Lk_mc_forward[]
132	vtbl.8	d0, {q12}, d6	@ vpshufb	%xmm3,	%xmm12,	%xmm0	# 0 = sb1t
133	vtbl.8	d1, {q12}, d7
134	veor	q4, q4, q5		@ vpxor		%xmm5,	%xmm4,	%xmm4	# 4 = sb1u + k
135	vtbl.8	d10, {q15}, d4	@ vpshufb	%xmm2,	%xmm15,	%xmm5	# 4 = sb2u
136	vtbl.8	d11, {q15}, d5
137	veor	q0, q0, q4		@ vpxor		%xmm4,	%xmm0,	%xmm0	# 0 = A
138	vtbl.8	d4, {q14}, d6	@ vpshufb	%xmm3,	%xmm14,	%xmm2	# 2 = sb2t
139	vtbl.8	d5, {q14}, d7
140	vld1.64	{q4}, [r10]		@ vmovdqa	(%r11,%r10), %xmm4	# .Lk_mc_backward[]
141	vtbl.8	d6, {q0}, d2	@ vpshufb	%xmm1,	%xmm0,	%xmm3	# 0 = B
142	vtbl.8	d7, {q0}, d3
143	veor	q2, q2, q5		@ vpxor		%xmm5,	%xmm2,	%xmm2	# 2 = 2A
144	@ Write to q5 instead of q0, so the table and destination registers do
145	@ not overlap.
146	vtbl.8	d10, {q0}, d8	@ vpshufb	%xmm4,	%xmm0,	%xmm0	# 3 = D
147	vtbl.8	d11, {q0}, d9
148	veor	q3, q3, q2		@ vpxor		%xmm2,	%xmm3,	%xmm3	# 0 = 2A+B
149	vtbl.8	d8, {q3}, d2	@ vpshufb	%xmm1,	%xmm3,	%xmm4	# 0 = 2B+C
150	vtbl.8	d9, {q3}, d3
151	@ Here we restore the original q0/q5 usage.
152	veor	q0, q5, q3		@ vpxor		%xmm3,	%xmm0,	%xmm0	# 3 = 2A+B+D
153	and	r11, r11, #~(1<<6)	@ and		$0x30,	%r11		# ... mod 4
154	veor	q0, q0, q4		@ vpxor		%xmm4,	%xmm0, %xmm0	# 0 = 2A+3B+C+D
155	subs	r8, r8, #1		@ nr--
156
157.Lenc_entry:
158	@ top of round
159	vand	q1, q0, q9		@ vpand		%xmm0,	%xmm9,	%xmm1   # 0 = k
160	vshr.u8	q0, q0, #4		@ vpsrlb	$4,	%xmm0,	%xmm0	# 1 = i
161	vtbl.8	d10, {q11}, d2	@ vpshufb	%xmm1,	%xmm11,	%xmm5	# 2 = a/k
162	vtbl.8	d11, {q11}, d3
163	veor	q1, q1, q0		@ vpxor		%xmm0,	%xmm1,	%xmm1	# 0 = j
164	vtbl.8	d6, {q10}, d0	@ vpshufb	%xmm0, 	%xmm10,	%xmm3  	# 3 = 1/i
165	vtbl.8	d7, {q10}, d1
166	vtbl.8	d8, {q10}, d2	@ vpshufb	%xmm1, 	%xmm10,	%xmm4  	# 4 = 1/j
167	vtbl.8	d9, {q10}, d3
168	veor	q3, q3, q5		@ vpxor		%xmm5,	%xmm3,	%xmm3	# 3 = iak = 1/i + a/k
169	veor	q4, q4, q5		@ vpxor		%xmm5,	%xmm4,	%xmm4  	# 4 = jak = 1/j + a/k
170	vtbl.8	d4, {q10}, d6	@ vpshufb	%xmm3,	%xmm10,	%xmm2  	# 2 = 1/iak
171	vtbl.8	d5, {q10}, d7
172	vtbl.8	d6, {q10}, d8	@ vpshufb	%xmm4,	%xmm10,	%xmm3	# 3 = 1/jak
173	vtbl.8	d7, {q10}, d9
174	veor	q2, q2, q1		@ vpxor		%xmm1,	%xmm2,	%xmm2  	# 2 = io
175	veor	q3, q3, q0		@ vpxor		%xmm0,	%xmm3,	%xmm3	# 3 = jo
176	vld1.64	{q5}, [r9]!		@ vmovdqu	(%r9),	%xmm5
177	bne	.Lenc_loop
178
179	@ middle of last round
180	add	r10, r11, #0x80
181
182	adr	r11, .Lk_sbo
183	@ Read to q1 instead of q4, so the vtbl.8 instruction below does not
184	@ overlap table and destination registers.
185	vld1.64	{q1}, [r11]!		@ vmovdqa	-0x60(%r10), %xmm4	# 3 : sbou
186	vld1.64	{q0}, [r11]		@ vmovdqa	-0x50(%r10), %xmm0	# 0 : sbot	.Lk_sbo+16
187	vtbl.8	d8, {q1}, d4	@ vpshufb	%xmm2,	%xmm4,	%xmm4	# 4 = sbou
188	vtbl.8	d9, {q1}, d5
189	vld1.64	{q1}, [r10]		@ vmovdqa	0x40(%r11,%r10), %xmm1	# .Lk_sr[]
190	@ Write to q2 instead of q0 below, to avoid overlapping table and
191	@ destination registers.
192	vtbl.8	d4, {q0}, d6	@ vpshufb	%xmm3,	%xmm0,	%xmm0	# 0 = sb1t
193	vtbl.8	d5, {q0}, d7
194	veor	q4, q4, q5		@ vpxor	%xmm5,	%xmm4,	%xmm4	# 4 = sb1u + k
195	veor	q2, q2, q4		@ vpxor	%xmm4,	%xmm0,	%xmm0	# 0 = A
196	@ Here we restore the original q0/q2 usage.
197	vtbl.8	d0, {q2}, d2	@ vpshufb	%xmm1,	%xmm0,	%xmm0
198	vtbl.8	d1, {q2}, d3
199	bx	lr
200.size	_vpaes_encrypt_core,.-_vpaes_encrypt_core
201
202.globl	vpaes_encrypt
203.hidden	vpaes_encrypt
204.type	vpaes_encrypt,%function
205.align	4
206vpaes_encrypt:
207	@ _vpaes_encrypt_core uses r8-r11. Round up to r7-r11 to maintain stack
208	@ alignment.
209	stmdb	sp!, {r7,r8,r9,r10,r11,lr}
210	@ _vpaes_encrypt_core uses q4-q5 (d8-d11), which are callee-saved.
211	vstmdb	sp!, {d8,d9,d10,d11}
212
213	vld1.64	{q0}, [r0]
214	bl	_vpaes_preheat
215	bl	_vpaes_encrypt_core
216	vst1.64	{q0}, [r1]
217
218	vldmia	sp!, {d8,d9,d10,d11}
219	ldmia	sp!, {r7,r8,r9,r10,r11, pc}	@ return
220.size	vpaes_encrypt,.-vpaes_encrypt
221@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
222@@                                                    @@
223@@                  AES key schedule                  @@
224@@                                                    @@
225@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
226
227@ This function diverges from both x86_64 and armv7 in which constants are
228@ pinned. x86_64 has a common preheat function for all operations. aarch64
229@ separates them because it has enough registers to pin nearly all constants.
230@ armv7 does not have enough registers, but needing explicit loads and stores
231@ also complicates using x86_64's register allocation directly.
232@
233@ We pin some constants for convenience and leave q14 and q15 free to load
234@ others on demand.
235
236@
237@  Key schedule constants
238@
239.type	_vpaes_key_consts,%object
240.align	4
241_vpaes_key_consts:
242.Lk_rcon:@ rcon
243.quad	0x1F8391B9AF9DEEB6, 0x702A98084D7C7D81
244
245.Lk_opt:@ output transform
246.quad	0xFF9F4929D6B66000, 0xF7974121DEBE6808
247.quad	0x01EDBD5150BCEC00, 0xE10D5DB1B05C0CE0
248.Lk_deskew:@ deskew tables: inverts the sbox's "skew"
249.quad	0x07E4A34047A4E300, 0x1DFEB95A5DBEF91A
250.quad	0x5F36B5DC83EA6900, 0x2841C2ABF49D1E77
251.size	_vpaes_key_consts,.-_vpaes_key_consts
252
253.type	_vpaes_key_preheat,%function
254.align	4
255_vpaes_key_preheat:
256	adr	r11, .Lk_rcon
257	vmov.i8	q12, #0x5b			@ .Lk_s63
258	adr	r10, .Lk_inv			@ Must be aligned to 8 mod 16.
259	vmov.i8	q9, #0x0f			@ .Lk_s0F
260	vld1.64	{q10,q11}, [r10]		@ .Lk_inv
261	vld1.64	{q8}, [r11]			@ .Lk_rcon
262	bx	lr
263.size	_vpaes_key_preheat,.-_vpaes_key_preheat
264
265.type	_vpaes_schedule_core,%function
266.align	4
267_vpaes_schedule_core:
268	@ We only need to save lr, but ARM requires an 8-byte stack alignment,
269	@ so save an extra register.
270	stmdb	sp!, {r3,lr}
271
272	bl	_vpaes_key_preheat	@ load the tables
273
274	adr	r11, .Lk_ipt		@ Must be aligned to 8 mod 16.
275	vld1.64	{q0}, [r0]!		@ vmovdqu	(%rdi),	%xmm0		# load key (unaligned)
276
277	@ input transform
278	@ Use q4 here rather than q3 so .Lschedule_am_decrypting does not
279	@ overlap table and destination.
280	vmov	q4, q0			@ vmovdqa	%xmm0,	%xmm3
281	bl	_vpaes_schedule_transform
282	adr	r10, .Lk_sr		@ Must be aligned to 8 mod 16.
283	vmov	q7, q0			@ vmovdqa	%xmm0,	%xmm7
284
285	add	r8, r8, r10
286
287	@ encrypting, output zeroth round key after transform
288	vst1.64	{q0}, [r2]		@ vmovdqu	%xmm0,	(%rdx)
289
290	@ *ring*: Decryption removed.
291
292.Lschedule_go:
293	cmp	r1, #192		@ cmp	$192,	%esi
294	bhi	.Lschedule_256
295	@ 128: fall though
296
297@@
298@@  .schedule_128
299@@
300@@  128-bit specific part of key schedule.
301@@
302@@  This schedule is really simple, because all its parts
303@@  are accomplished by the subroutines.
304@@
305.Lschedule_128:
306	mov	r0, #10		@ mov	$10, %esi
307
308.Loop_schedule_128:
309	bl	_vpaes_schedule_round
310	subs	r0, r0, #1		@ dec	%esi
311	beq	.Lschedule_mangle_last
312	bl	_vpaes_schedule_mangle	@ write output
313	b	.Loop_schedule_128
314
315@@
316@@  .aes_schedule_256
317@@
318@@  256-bit specific part of key schedule.
319@@
320@@  The structure here is very similar to the 128-bit
321@@  schedule, but with an additional "low side" in
322@@  q6.  The low side's rounds are the same as the
323@@  high side's, except no rcon and no rotation.
324@@
325.align	4
326.Lschedule_256:
327	vld1.64	{q0}, [r0]			@ vmovdqu	16(%rdi),%xmm0		# load key part 2 (unaligned)
328	bl	_vpaes_schedule_transform	@ input transform
329	mov	r0, #7			@ mov	$7, %esi
330
331.Loop_schedule_256:
332	bl	_vpaes_schedule_mangle		@ output low result
333	vmov	q6, q0				@ vmovdqa	%xmm0,	%xmm6		# save cur_lo in xmm6
334
335	@ high round
336	bl	_vpaes_schedule_round
337	subs	r0, r0, #1			@ dec	%esi
338	beq	.Lschedule_mangle_last
339	bl	_vpaes_schedule_mangle
340
341	@ low round. swap xmm7 and xmm6
342	vdup.32	q0, d1[1]		@ vpshufd	$0xFF,	%xmm0,	%xmm0
343	vmov.i8	q4, #0
344	vmov	q5, q7			@ vmovdqa	%xmm7,	%xmm5
345	vmov	q7, q6			@ vmovdqa	%xmm6,	%xmm7
346	bl	_vpaes_schedule_low_round
347	vmov	q7, q5			@ vmovdqa	%xmm5,	%xmm7
348
349	b	.Loop_schedule_256
350
351@@
352@@  .aes_schedule_mangle_last
353@@
354@@  Mangler for last round of key schedule
355@@  Mangles q0
356@@    when encrypting, outputs out(q0) ^ 63
357@@    when decrypting, outputs unskew(q0)
358@@
359@@  Always called right before return... jumps to cleanup and exits
360@@
361.align	4
362.Lschedule_mangle_last:
363	@ schedule last round key from xmm0
364	adr	r11, .Lk_deskew			@ lea	.Lk_deskew(%rip),%r11	# prepare to deskew
365
366	@ encrypting
367	vld1.64	{q1}, [r8]		@ vmovdqa	(%r8,%r10),%xmm1
368	adr	r11, .Lk_opt		@ lea		.Lk_opt(%rip),	%r11		# prepare to output transform
369	add	r2, r2, #32		@ add		$32,	%rdx
370	vmov	q2, q0
371	vtbl.8	d0, {q2}, d2	@ vpshufb	%xmm1,	%xmm0,	%xmm0		# output permute
372	vtbl.8	d1, {q2}, d3
373
374.Lschedule_mangle_last_dec:
375	sub	r2, r2, #16			@ add	$-16,	%rdx
376	veor	q0, q0, q12			@ vpxor	.Lk_s63(%rip),	%xmm0,	%xmm0
377	bl	_vpaes_schedule_transform	@ output transform
378	vst1.64	{q0}, [r2]			@ vmovdqu	%xmm0,	(%rdx)		# save last key
379
380	@ cleanup
381	veor	q0, q0, q0		@ vpxor	%xmm0,	%xmm0,	%xmm0
382	veor	q1, q1, q1		@ vpxor	%xmm1,	%xmm1,	%xmm1
383	veor	q2, q2, q2		@ vpxor	%xmm2,	%xmm2,	%xmm2
384	veor	q3, q3, q3		@ vpxor	%xmm3,	%xmm3,	%xmm3
385	veor	q4, q4, q4		@ vpxor	%xmm4,	%xmm4,	%xmm4
386	veor	q5, q5, q5		@ vpxor	%xmm5,	%xmm5,	%xmm5
387	veor	q6, q6, q6		@ vpxor	%xmm6,	%xmm6,	%xmm6
388	veor	q7, q7, q7		@ vpxor	%xmm7,	%xmm7,	%xmm7
389	ldmia	sp!, {r3,pc}		@ return
390.size	_vpaes_schedule_core,.-_vpaes_schedule_core
391
392@@
393@@  .aes_schedule_round
394@@
395@@  Runs one main round of the key schedule on q0, q7
396@@
397@@  Specifically, runs subbytes on the high dword of q0
398@@  then rotates it by one byte and xors into the low dword of
399@@  q7.
400@@
401@@  Adds rcon from low byte of q8, then rotates q8 for
402@@  next rcon.
403@@
404@@  Smears the dwords of q7 by xoring the low into the
405@@  second low, result into third, result into highest.
406@@
407@@  Returns results in q7 = q0.
408@@  Clobbers q1-q4, r11.
409@@
410.type	_vpaes_schedule_round,%function
411.align	4
412_vpaes_schedule_round:
413	@ extract rcon from xmm8
414	vmov.i8	q4, #0				@ vpxor		%xmm4,	%xmm4,	%xmm4
415	vext.8	q1, q8, q4, #15		@ vpalignr	$15,	%xmm8,	%xmm4,	%xmm1
416	vext.8	q8, q8, q8, #15	@ vpalignr	$15,	%xmm8,	%xmm8,	%xmm8
417	veor	q7, q7, q1			@ vpxor		%xmm1,	%xmm7,	%xmm7
418
419	@ rotate
420	vdup.32	q0, d1[1]			@ vpshufd	$0xFF,	%xmm0,	%xmm0
421	vext.8	q0, q0, q0, #1			@ vpalignr	$1,	%xmm0,	%xmm0,	%xmm0
422
423	@ fall through...
424
425	@ low round: same as high round, but no rotation and no rcon.
426_vpaes_schedule_low_round:
427	@ The x86_64 version pins .Lk_sb1 in %xmm13 and .Lk_sb1+16 in %xmm12.
428	@ We pin other values in _vpaes_key_preheat, so load them now.
429	adr	r11, .Lk_sb1
430	vld1.64	{q14,q15}, [r11]
431
432	@ smear xmm7
433	vext.8	q1, q4, q7, #12			@ vpslldq	$4,	%xmm7,	%xmm1
434	veor	q7, q7, q1			@ vpxor	%xmm1,	%xmm7,	%xmm7
435	vext.8	q4, q4, q7, #8			@ vpslldq	$8,	%xmm7,	%xmm4
436
437	@ subbytes
438	vand	q1, q0, q9			@ vpand		%xmm9,	%xmm0,	%xmm1		# 0 = k
439	vshr.u8	q0, q0, #4			@ vpsrlb	$4,	%xmm0,	%xmm0		# 1 = i
440	veor	q7, q7, q4			@ vpxor		%xmm4,	%xmm7,	%xmm7
441	vtbl.8	d4, {q11}, d2		@ vpshufb	%xmm1,	%xmm11,	%xmm2		# 2 = a/k
442	vtbl.8	d5, {q11}, d3
443	veor	q1, q1, q0			@ vpxor		%xmm0,	%xmm1,	%xmm1		# 0 = j
444	vtbl.8	d6, {q10}, d0		@ vpshufb	%xmm0, 	%xmm10,	%xmm3		# 3 = 1/i
445	vtbl.8	d7, {q10}, d1
446	veor	q3, q3, q2			@ vpxor		%xmm2,	%xmm3,	%xmm3		# 3 = iak = 1/i + a/k
447	vtbl.8	d8, {q10}, d2		@ vpshufb	%xmm1,	%xmm10,	%xmm4		# 4 = 1/j
448	vtbl.8	d9, {q10}, d3
449	veor	q7, q7, q12			@ vpxor		.Lk_s63(%rip),	%xmm7,	%xmm7
450	vtbl.8	d6, {q10}, d6		@ vpshufb	%xmm3,	%xmm10,	%xmm3		# 2 = 1/iak
451	vtbl.8	d7, {q10}, d7
452	veor	q4, q4, q2			@ vpxor		%xmm2,	%xmm4,	%xmm4		# 4 = jak = 1/j + a/k
453	vtbl.8	d4, {q10}, d8		@ vpshufb	%xmm4,	%xmm10,	%xmm2		# 3 = 1/jak
454	vtbl.8	d5, {q10}, d9
455	veor	q3, q3, q1			@ vpxor		%xmm1,	%xmm3,	%xmm3		# 2 = io
456	veor	q2, q2, q0			@ vpxor		%xmm0,	%xmm2,	%xmm2		# 3 = jo
457	vtbl.8	d8, {q15}, d6		@ vpshufb	%xmm3,	%xmm13,	%xmm4		# 4 = sbou
458	vtbl.8	d9, {q15}, d7
459	vtbl.8	d2, {q14}, d4		@ vpshufb	%xmm2,	%xmm12,	%xmm1		# 0 = sb1t
460	vtbl.8	d3, {q14}, d5
461	veor	q1, q1, q4			@ vpxor		%xmm4,	%xmm1,	%xmm1		# 0 = sbox output
462
463	@ add in smeared stuff
464	veor	q0, q1, q7			@ vpxor	%xmm7,	%xmm1,	%xmm0
465	veor	q7, q1, q7			@ vmovdqa	%xmm0,	%xmm7
466	bx	lr
467.size	_vpaes_schedule_round,.-_vpaes_schedule_round
468
469@@
470@@  .aes_schedule_transform
471@@
472@@  Linear-transform q0 according to tables at [r11]
473@@
474@@  Requires that q9 = 0x0F0F... as in preheat
475@@  Output in q0
476@@  Clobbers q1, q2, q14, q15
477@@
478.type	_vpaes_schedule_transform,%function
479.align	4
480_vpaes_schedule_transform:
481	vld1.64	{q14,q15}, [r11]	@ vmovdqa	(%r11),	%xmm2 	# lo
482					@ vmovdqa	16(%r11),	%xmm1 # hi
483	vand	q1, q0, q9		@ vpand	%xmm9,	%xmm0,	%xmm1
484	vshr.u8	q0, q0, #4		@ vpsrlb	$4,	%xmm0,	%xmm0
485	vtbl.8	d4, {q14}, d2	@ vpshufb	%xmm1,	%xmm2,	%xmm2
486	vtbl.8	d5, {q14}, d3
487	vtbl.8	d0, {q15}, d0	@ vpshufb	%xmm0,	%xmm1,	%xmm0
488	vtbl.8	d1, {q15}, d1
489	veor	q0, q0, q2		@ vpxor	%xmm2,	%xmm0,	%xmm0
490	bx	lr
491.size	_vpaes_schedule_transform,.-_vpaes_schedule_transform
492
493@@
494@@  .aes_schedule_mangle
495@@
496@@  Mangles q0 from (basis-transformed) standard version
497@@  to our version.
498@@
499@@  On encrypt,
500@@    xor with 0x63
501@@    multiply by circulant 0,1,1,1
502@@    apply shiftrows transform
503@@
504@@  On decrypt,
505@@    xor with 0x63
506@@    multiply by "inverse mixcolumns" circulant E,B,D,9
507@@    deskew
508@@    apply shiftrows transform
509@@
510@@
511@@  Writes out to [r2], and increments or decrements it
512@@  Keeps track of round number mod 4 in r8
513@@  Preserves q0
514@@  Clobbers q1-q5
515@@
516.type	_vpaes_schedule_mangle,%function
517.align	4
518_vpaes_schedule_mangle:
519	tst	r3, r3
520	vmov	q4, q0			@ vmovdqa	%xmm0,	%xmm4	# save xmm0 for later
521	adr	r11, .Lk_mc_forward	@ Must be aligned to 8 mod 16.
522	vld1.64	{q5}, [r11]		@ vmovdqa	.Lk_mc_forward(%rip),%xmm5
523
524	@ encrypting
525	@ Write to q2 so we do not overlap table and destination below.
526	veor	q2, q0, q12		@ vpxor		.Lk_s63(%rip),	%xmm0,	%xmm4
527	add	r2, r2, #16		@ add		$16,	%rdx
528	vtbl.8	d8, {q2}, d10	@ vpshufb	%xmm5,	%xmm4,	%xmm4
529	vtbl.8	d9, {q2}, d11
530	vtbl.8	d2, {q4}, d10	@ vpshufb	%xmm5,	%xmm4,	%xmm1
531	vtbl.8	d3, {q4}, d11
532	vtbl.8	d6, {q1}, d10	@ vpshufb	%xmm5,	%xmm1,	%xmm3
533	vtbl.8	d7, {q1}, d11
534	veor	q4, q4, q1		@ vpxor		%xmm1,	%xmm4,	%xmm4
535	vld1.64	{q1}, [r8]		@ vmovdqa	(%r8,%r10),	%xmm1
536	veor	q3, q3, q4		@ vpxor		%xmm4,	%xmm3,	%xmm3
537
538.Lschedule_mangle_both:
539	@ Write to q2 so table and destination do not overlap.
540	vtbl.8	d4, {q3}, d2	@ vpshufb	%xmm1,	%xmm3,	%xmm3
541	vtbl.8	d5, {q3}, d3
542	add	r8, r8, #64-16		@ add	$-16,	%r8
543	and	r8, r8, #~(1<<6)	@ and	$0x30,	%r8
544	vst1.64	{q2}, [r2]		@ vmovdqu	%xmm3,	(%rdx)
545	bx	lr
546.size	_vpaes_schedule_mangle,.-_vpaes_schedule_mangle
547
548.globl	vpaes_set_encrypt_key
549.hidden	vpaes_set_encrypt_key
550.type	vpaes_set_encrypt_key,%function
551.align	4
552vpaes_set_encrypt_key:
553	stmdb	sp!, {r7,r8,r9,r10,r11, lr}
554	vstmdb	sp!, {d8,d9,d10,d11,d12,d13,d14,d15}
555
556	lsr	r9, r1, #5		@ shr	$5,%eax
557	add	r9, r9, #5		@ $5,%eax
558	str	r9, [r2,#240]		@ mov	%eax,240(%rdx)	# AES_KEY->rounds = nbits/32+5;
559
560	mov	r3, #0		@ mov	$0,%ecx
561	mov	r8, #0x30		@ mov	$0x30,%r8d
562	bl	_vpaes_schedule_core
563	eor	r0, r0, r0
564
565	vldmia	sp!, {d8,d9,d10,d11,d12,d13,d14,d15}
566	ldmia	sp!, {r7,r8,r9,r10,r11, pc}	@ return
567.size	vpaes_set_encrypt_key,.-vpaes_set_encrypt_key
568
569@ Additional constants for converting to bsaes.
570.type	_vpaes_convert_consts,%object
571.align	4
572_vpaes_convert_consts:
573@ .Lk_opt_then_skew applies skew(opt(x)) XOR 0x63, where skew is the linear
574@ transform in the AES S-box. 0x63 is incorporated into the low half of the
575@ table. This was computed with the following script:
576@
577@   def u64s_to_u128(x, y):
578@       return x | (y << 64)
579@   def u128_to_u64s(w):
580@       return w & ((1<<64)-1), w >> 64
581@   def get_byte(w, i):
582@       return (w >> (i*8)) & 0xff
583@   def apply_table(table, b):
584@       lo = b & 0xf
585@       hi = b >> 4
586@       return get_byte(table[0], lo) ^ get_byte(table[1], hi)
587@   def opt(b):
588@       table = [
589@           u64s_to_u128(0xFF9F4929D6B66000, 0xF7974121DEBE6808),
590@           u64s_to_u128(0x01EDBD5150BCEC00, 0xE10D5DB1B05C0CE0),
591@       ]
592@       return apply_table(table, b)
593@   def rot_byte(b, n):
594@       return 0xff & ((b << n) | (b >> (8-n)))
595@   def skew(x):
596@       return (x ^ rot_byte(x, 1) ^ rot_byte(x, 2) ^ rot_byte(x, 3) ^
597@               rot_byte(x, 4))
598@   table = [0, 0]
599@   for i in range(16):
600@       table[0] |= (skew(opt(i)) ^ 0x63) << (i*8)
601@       table[1] |= skew(opt(i<<4)) << (i*8)
602@   print("	.quad	0x%016x, 0x%016x" % u128_to_u64s(table[0]))
603@   print("	.quad	0x%016x, 0x%016x" % u128_to_u64s(table[1]))
604.Lk_opt_then_skew:
605.quad	0x9cb8436798bc4763, 0x6440bb9f6044bf9b
606.quad	0x1f30062936192f00, 0xb49bad829db284ab
607
608@ void vpaes_encrypt_key_to_bsaes(AES_KEY *bsaes, const AES_KEY *vpaes);
609.globl	vpaes_encrypt_key_to_bsaes
610.hidden	vpaes_encrypt_key_to_bsaes
611.type	vpaes_encrypt_key_to_bsaes,%function
612.align	4
613vpaes_encrypt_key_to_bsaes:
614	stmdb	sp!, {r11, lr}
615
616	@ See _vpaes_schedule_core for the key schedule logic. In particular,
617	@ _vpaes_schedule_transform(.Lk_ipt) (section 2.2 of the paper),
618	@ _vpaes_schedule_mangle (section 4.3), and .Lschedule_mangle_last
619	@ contain the transformations not in the bsaes representation. This
620	@ function inverts those transforms.
621	@
622	@ Note also that bsaes-armv7.pl expects aes-armv4.pl's key
623	@ representation, which does not match the other aes_nohw_*
624	@ implementations. The ARM aes_nohw_* stores each 32-bit word
625	@ byteswapped, as a convenience for (unsupported) big-endian ARM, at the
626	@ cost of extra REV and VREV32 operations in little-endian ARM.
627
628	vmov.i8	q9, #0x0f		@ Required by _vpaes_schedule_transform
629	adr	r2, .Lk_mc_forward	@ Must be aligned to 8 mod 16.
630	add	r3, r2, 0x90		@ .Lk_sr+0x10-.Lk_mc_forward = 0x90 (Apple's toolchain doesn't support the expression)
631
632	vld1.64	{q12}, [r2]
633	vmov.i8	q10, #0x5b		@ .Lk_s63 from vpaes-x86_64
634	adr	r11, .Lk_opt		@ Must be aligned to 8 mod 16.
635	vmov.i8	q11, #0x63		@ .LK_s63 without .Lk_ipt applied
636
637	@ vpaes stores one fewer round count than bsaes, but the number of keys
638	@ is the same.
639	ldr	r2, [r1,#240]
640	add	r2, r2, #1
641	str	r2, [r0,#240]
642
643	@ The first key is transformed with _vpaes_schedule_transform(.Lk_ipt).
644	@ Invert this with .Lk_opt.
645	vld1.64	{q0}, [r1]!
646	bl	_vpaes_schedule_transform
647	vrev32.8	q0, q0
648	vst1.64	{q0}, [r0]!
649
650	@ The middle keys have _vpaes_schedule_transform(.Lk_ipt) applied,
651	@ followed by _vpaes_schedule_mangle. _vpaes_schedule_mangle XORs 0x63,
652	@ multiplies by the circulant 0,1,1,1, then applies ShiftRows.
653.Loop_enc_key_to_bsaes:
654	vld1.64	{q0}, [r1]!
655
656	@ Invert the ShiftRows step (see .Lschedule_mangle_both). Note we cycle
657	@ r3 in the opposite direction and start at .Lk_sr+0x10 instead of 0x30.
658	@ We use r3 rather than r8 to avoid a callee-saved register.
659	vld1.64	{q1}, [r3]
660	vtbl.8	d4, {q0}, d2
661	vtbl.8	d5, {q0}, d3
662	add	r3, r3, #16
663	and	r3, r3, #~(1<<6)
664	vmov	q0, q2
665
666	@ Handle the last key differently.
667	subs	r2, r2, #1
668	beq	.Loop_enc_key_to_bsaes_last
669
670	@ Multiply by the circulant. This is its own inverse.
671	vtbl.8	d2, {q0}, d24
672	vtbl.8	d3, {q0}, d25
673	vmov	q0, q1
674	vtbl.8	d4, {q1}, d24
675	vtbl.8	d5, {q1}, d25
676	veor	q0, q0, q2
677	vtbl.8	d2, {q2}, d24
678	vtbl.8	d3, {q2}, d25
679	veor	q0, q0, q1
680
681	@ XOR and finish.
682	veor	q0, q0, q10
683	bl	_vpaes_schedule_transform
684	vrev32.8	q0, q0
685	vst1.64	{q0}, [r0]!
686	b	.Loop_enc_key_to_bsaes
687
688.Loop_enc_key_to_bsaes_last:
689	@ The final key does not have a basis transform (note
690	@ .Lschedule_mangle_last inverts the original transform). It only XORs
691	@ 0x63 and applies ShiftRows. The latter was already inverted in the
692	@ loop. Note that, because we act on the original representation, we use
693	@ q11, not q10.
694	veor	q0, q0, q11
695	vrev32.8	q0, q0
696	vst1.64	{q0}, [r0]
697
698	@ Wipe registers which contained key material.
699	veor	q0, q0, q0
700	veor	q1, q1, q1
701	veor	q2, q2, q2
702
703	ldmia	sp!, {r11, pc}	@ return
704.size	vpaes_encrypt_key_to_bsaes,.-vpaes_encrypt_key_to_bsaes
705.globl	vpaes_ctr32_encrypt_blocks
706.hidden	vpaes_ctr32_encrypt_blocks
707.type	vpaes_ctr32_encrypt_blocks,%function
708.align	4
709vpaes_ctr32_encrypt_blocks:
710	mov	ip, sp
711	stmdb	sp!, {r7,r8,r9,r10,r11, lr}
712	@ This function uses q4-q7 (d8-d15), which are callee-saved.
713	vstmdb	sp!, {d8,d9,d10,d11,d12,d13,d14,d15}
714
715	cmp	r2, #0
716	@ r8 is passed on the stack.
717	ldr	r8, [ip]
718	beq	.Lctr32_done
719
720	@ _vpaes_encrypt_core expects the key in r2, so swap r2 and r3.
721	mov	r9, r3
722	mov	r3, r2
723	mov	r2, r9
724
725	@ Load the IV and counter portion.
726	ldr	r7, [r8, #12]
727	vld1.8	{q7}, [r8]
728
729	bl	_vpaes_preheat
730	rev	r7, r7		@ The counter is big-endian.
731
732.Lctr32_loop:
733	vmov	q0, q7
734	vld1.8	{q6}, [r0]!		@ .Load input ahead of time
735	bl	_vpaes_encrypt_core
736	veor	q0, q0, q6		@ XOR input and result
737	vst1.8	{q0}, [r1]!
738	subs	r3, r3, #1
739	@ Update the counter.
740	add	r7, r7, #1
741	rev	r9, r7
742	vmov.32	d15[1], r9
743	bne	.Lctr32_loop
744
745.Lctr32_done:
746	vldmia	sp!, {d8,d9,d10,d11,d12,d13,d14,d15}
747	ldmia	sp!, {r7,r8,r9,r10,r11, pc}	@ return
748.size	vpaes_ctr32_encrypt_blocks,.-vpaes_ctr32_encrypt_blocks
749#endif
750#endif  // !OPENSSL_NO_ASM
751.section	.note.GNU-stack,"",%progbits
752