external/boringssl: Sync to 3a3552247ecb0bfb260a36d9da7a3bce7fdc3f8a.

This includes the following changes:

https://boringssl.googlesource.com/boringssl/+log/e60b080dda138e1dd02d99fb34641ac22e46c85d..3a3552247ecb0bfb260a36d9da7a3bce7fdc3f8a

Also changes Android.bp to compile with execute-only memory again as
this should no longer be necessary with the mprotect changes in
https://boringssl.googlesource.com/boringssl/+/09a9ec036030ac84896f5143548d05f3951d1817

Bug: 134580074
Test: atest CtsLibcoreTestCases CtsLibcoreOkHttpTestCases
Change-Id: I0ec54998afd2e0b40ec930716397e20aa3c21bf2

1 files changed, 1234 insertions, 0 deletions

diff --git a/linux-arm/crypto/fipsmodule/vpaes-armv7.S b/linux-arm/crypto/fipsmodule/vpaes-armv7.S
new file mode 100644
index 00000000..a9243665
--- /dev/null
+++ b/linux-arm/crypto/fipsmodule/vpaes-armv7.S
@@ -0,0 +1,1234 @@
+// This file is generated from a similarly-named Perl script in the BoringSSL
+// source tree. Do not edit by hand.
+
+#if !defined(__has_feature)
+#define __has_feature(x) 0
+#endif
+#if __has_feature(memory_sanitizer) && !defined(OPENSSL_NO_ASM)
+#define OPENSSL_NO_ASM
+#endif
+
+#if !defined(OPENSSL_NO_ASM)
+#if defined(__arm__)
+#if defined(BORINGSSL_PREFIX)
+#include <boringssl_prefix_symbols_asm.h>
+#endif
+.syntax	unified
+
+.arch	armv7-a
+.fpu	neon
+
+#if defined(__thumb2__)
+.thumb
+#else
+.code	32
+#endif
+
+.text
+
+.type	_vpaes_consts,%object
+.align	7	@ totally strategic alignment
+_vpaes_consts:
+.Lk_mc_forward:@ mc_forward
+.quad	0x0407060500030201, 0x0C0F0E0D080B0A09
+.quad	0x080B0A0904070605, 0x000302010C0F0E0D
+.quad	0x0C0F0E0D080B0A09, 0x0407060500030201
+.quad	0x000302010C0F0E0D, 0x080B0A0904070605
+.Lk_mc_backward:@ mc_backward
+.quad	0x0605040702010003, 0x0E0D0C0F0A09080B
+.quad	0x020100030E0D0C0F, 0x0A09080B06050407
+.quad	0x0E0D0C0F0A09080B, 0x0605040702010003
+.quad	0x0A09080B06050407, 0x020100030E0D0C0F
+.Lk_sr:@ sr
+.quad	0x0706050403020100, 0x0F0E0D0C0B0A0908
+.quad	0x030E09040F0A0500, 0x0B06010C07020D08
+.quad	0x0F060D040B020900, 0x070E050C030A0108
+.quad	0x0B0E0104070A0D00, 0x0306090C0F020508
+
+@
+@ "Hot" constants
+@
+.Lk_inv:@ inv, inva
+.quad	0x0E05060F0D080180, 0x040703090A0B0C02
+.quad	0x01040A060F0B0780, 0x030D0E0C02050809
+.Lk_ipt:@ input transform (lo, hi)
+.quad	0xC2B2E8985A2A7000, 0xCABAE09052227808
+.quad	0x4C01307D317C4D00, 0xCD80B1FCB0FDCC81
+.Lk_sbo:@ sbou, sbot
+.quad	0xD0D26D176FBDC700, 0x15AABF7AC502A878
+.quad	0xCFE474A55FBB6A00, 0x8E1E90D1412B35FA
+.Lk_sb1:@ sb1u, sb1t
+.quad	0x3618D415FAE22300, 0x3BF7CCC10D2ED9EF
+.quad	0xB19BE18FCB503E00, 0xA5DF7A6E142AF544
+.Lk_sb2:@ sb2u, sb2t
+.quad	0x69EB88400AE12900, 0xC2A163C8AB82234A
+.quad	0xE27A93C60B712400, 0x5EB7E955BC982FCD
+
+.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
+.align	2
+.size	_vpaes_consts,.-_vpaes_consts
+.align	6
+@@
+@@  _aes_preheat
+@@
+@@  Fills q9-q15 as specified below.
+@@
+.type	_vpaes_preheat,%function
+.align	4
+_vpaes_preheat:
+	adr	r10, .Lk_inv
+	vmov.i8	q9, #0x0f		@ .Lk_s0F
+	vld1.64	{q10,q11}, [r10]!	@ .Lk_inv
+	add	r10, r10, #64		@ Skip .Lk_ipt, .Lk_sbo
+	vld1.64	{q12,q13}, [r10]!	@ .Lk_sb1
+	vld1.64	{q14,q15}, [r10]	@ .Lk_sb2
+	bx	lr
+
+@@
+@@  _aes_encrypt_core
+@@
+@@  AES-encrypt q0.
+@@
+@@  Inputs:
+@@     q0 = input
+@@     q9-q15 as in _vpaes_preheat
+@@    [r2] = scheduled keys
+@@
+@@  Output in q0
+@@  Clobbers  q1-q5, r8-r11
+@@  Preserves q6-q8 so you get some local vectors
+@@
+@@
+.type	_vpaes_encrypt_core,%function
+.align	4
+_vpaes_encrypt_core:
+	mov	r9, r2
+	ldr	r8, [r2,#240]		@ pull rounds
+	adr	r11, .Lk_ipt
+	@ vmovdqa	.Lk_ipt(%rip),	%xmm2	# iptlo
+	@ vmovdqa	.Lk_ipt+16(%rip), %xmm3	# ipthi
+	vld1.64	{q2, q3}, [r11]
+	adr	r11, .Lk_mc_forward+16
+	vld1.64	{q5}, [r9]!		@ vmovdqu	(%r9),	%xmm5		# round0 key
+	vand	q1, q0, q9		@ vpand	%xmm9,	%xmm0,	%xmm1
+	vshr.u8	q0, q0, #4		@ vpsrlb	$4,	%xmm0,	%xmm0
+	vtbl.8	d2, {q2}, d2	@ vpshufb	%xmm1,	%xmm2,	%xmm1
+	vtbl.8	d3, {q2}, d3
+	vtbl.8	d4, {q3}, d0	@ vpshufb	%xmm0,	%xmm3,	%xmm2
+	vtbl.8	d5, {q3}, d1
+	veor	q0, q1, q5		@ vpxor	%xmm5,	%xmm1,	%xmm0
+	veor	q0, q0, q2		@ vpxor	%xmm2,	%xmm0,	%xmm0
+
+	@ .Lenc_entry ends with a bnz instruction which is normally paired with
+	@ subs in .Lenc_loop.
+	tst	r8, r8
+	b	.Lenc_entry
+
+.align	4
+.Lenc_loop:
+	@ middle of middle round
+	add	r10, r11, #0x40
+	vtbl.8	d8, {q13}, d4	@ vpshufb	%xmm2,	%xmm13,	%xmm4	# 4 = sb1u
+	vtbl.8	d9, {q13}, d5
+	vld1.64	{q1}, [r11]!		@ vmovdqa	-0x40(%r11,%r10), %xmm1	# .Lk_mc_forward[]
+	vtbl.8	d0, {q12}, d6	@ vpshufb	%xmm3,	%xmm12,	%xmm0	# 0 = sb1t
+	vtbl.8	d1, {q12}, d7
+	veor	q4, q4, q5		@ vpxor		%xmm5,	%xmm4,	%xmm4	# 4 = sb1u + k
+	vtbl.8	d10, {q15}, d4	@ vpshufb	%xmm2,	%xmm15,	%xmm5	# 4 = sb2u
+	vtbl.8	d11, {q15}, d5
+	veor	q0, q0, q4		@ vpxor		%xmm4,	%xmm0,	%xmm0	# 0 = A
+	vtbl.8	d4, {q14}, d6	@ vpshufb	%xmm3,	%xmm14,	%xmm2	# 2 = sb2t
+	vtbl.8	d5, {q14}, d7
+	vld1.64	{q4}, [r10]		@ vmovdqa	(%r11,%r10), %xmm4	# .Lk_mc_backward[]
+	vtbl.8	d6, {q0}, d2	@ vpshufb	%xmm1,	%xmm0,	%xmm3	# 0 = B
+	vtbl.8	d7, {q0}, d3
+	veor	q2, q2, q5		@ vpxor		%xmm5,	%xmm2,	%xmm2	# 2 = 2A
+	@ Write to q5 instead of q0, so the table and destination registers do
+	@ not overlap.
+	vtbl.8	d10, {q0}, d8	@ vpshufb	%xmm4,	%xmm0,	%xmm0	# 3 = D
+	vtbl.8	d11, {q0}, d9
+	veor	q3, q3, q2		@ vpxor		%xmm2,	%xmm3,	%xmm3	# 0 = 2A+B
+	vtbl.8	d8, {q3}, d2	@ vpshufb	%xmm1,	%xmm3,	%xmm4	# 0 = 2B+C
+	vtbl.8	d9, {q3}, d3
+	@ Here we restore the original q0/q5 usage.
+	veor	q0, q5, q3		@ vpxor		%xmm3,	%xmm0,	%xmm0	# 3 = 2A+B+D
+	and	r11, r11, #~(1<<6)	@ and		$0x30,	%r11		# ... mod 4
+	veor	q0, q0, q4		@ vpxor		%xmm4,	%xmm0, %xmm0	# 0 = 2A+3B+C+D
+	subs	r8, r8, #1		@ nr--
+
+.Lenc_entry:
+	@ top of round
+	vand	q1, q0, q9		@ vpand		%xmm0,	%xmm9,	%xmm1   # 0 = k
+	vshr.u8	q0, q0, #4		@ vpsrlb	$4,	%xmm0,	%xmm0	# 1 = i
+	vtbl.8	d10, {q11}, d2	@ vpshufb	%xmm1,	%xmm11,	%xmm5	# 2 = a/k
+	vtbl.8	d11, {q11}, d3
+	veor	q1, q1, q0		@ vpxor		%xmm0,	%xmm1,	%xmm1	# 0 = j
+	vtbl.8	d6, {q10}, d0	@ vpshufb	%xmm0, 	%xmm10,	%xmm3  	# 3 = 1/i
+	vtbl.8	d7, {q10}, d1
+	vtbl.8	d8, {q10}, d2	@ vpshufb	%xmm1, 	%xmm10,	%xmm4  	# 4 = 1/j
+	vtbl.8	d9, {q10}, d3
+	veor	q3, q3, q5		@ vpxor		%xmm5,	%xmm3,	%xmm3	# 3 = iak = 1/i + a/k
+	veor	q4, q4, q5		@ vpxor		%xmm5,	%xmm4,	%xmm4  	# 4 = jak = 1/j + a/k
+	vtbl.8	d4, {q10}, d6	@ vpshufb	%xmm3,	%xmm10,	%xmm2  	# 2 = 1/iak
+	vtbl.8	d5, {q10}, d7
+	vtbl.8	d6, {q10}, d8	@ vpshufb	%xmm4,	%xmm10,	%xmm3	# 3 = 1/jak
+	vtbl.8	d7, {q10}, d9
+	veor	q2, q2, q1		@ vpxor		%xmm1,	%xmm2,	%xmm2  	# 2 = io
+	veor	q3, q3, q0		@ vpxor		%xmm0,	%xmm3,	%xmm3	# 3 = jo
+	vld1.64	{q5}, [r9]!		@ vmovdqu	(%r9),	%xmm5
+	bne	.Lenc_loop
+
+	@ middle of last round
+	add	r10, r11, #0x80
+
+	adr	r11, .Lk_sbo
+	@ Read to q1 instead of q4, so the vtbl.8 instruction below does not
+	@ overlap table and destination registers.
+	vld1.64	{q1}, [r11]!		@ vmovdqa	-0x60(%r10), %xmm4	# 3 : sbou
+	vld1.64	{q0}, [r11]		@ vmovdqa	-0x50(%r10), %xmm0	# 0 : sbot	.Lk_sbo+16
+	vtbl.8	d8, {q1}, d4	@ vpshufb	%xmm2,	%xmm4,	%xmm4	# 4 = sbou
+	vtbl.8	d9, {q1}, d5
+	vld1.64	{q1}, [r10]		@ vmovdqa	0x40(%r11,%r10), %xmm1	# .Lk_sr[]
+	@ Write to q2 instead of q0 below, to avoid overlapping table and
+	@ destination registers.
+	vtbl.8	d4, {q0}, d6	@ vpshufb	%xmm3,	%xmm0,	%xmm0	# 0 = sb1t
+	vtbl.8	d5, {q0}, d7
+	veor	q4, q4, q5		@ vpxor	%xmm5,	%xmm4,	%xmm4	# 4 = sb1u + k
+	veor	q2, q2, q4		@ vpxor	%xmm4,	%xmm0,	%xmm0	# 0 = A
+	@ Here we restore the original q0/q2 usage.
+	vtbl.8	d0, {q2}, d2	@ vpshufb	%xmm1,	%xmm0,	%xmm0
+	vtbl.8	d1, {q2}, d3
+	bx	lr
+.size	_vpaes_encrypt_core,.-_vpaes_encrypt_core
+
+.globl	vpaes_encrypt
+.hidden	vpaes_encrypt
+.type	vpaes_encrypt,%function
+.align	4
+vpaes_encrypt:
+	@ _vpaes_encrypt_core uses r8-r11. Round up to r7-r11 to maintain stack
+	@ alignment.
+	stmdb	sp!, {r7,r8,r9,r10,r11,lr}
+	@ _vpaes_encrypt_core uses q4-q5 (d8-d11), which are callee-saved.
+	vstmdb	sp!, {d8,d9,d10,d11}
+
+	vld1.64	{q0}, [r0]
+	bl	_vpaes_preheat
+	bl	_vpaes_encrypt_core
+	vst1.64	{q0}, [r1]
+
+	vldmia	sp!, {d8,d9,d10,d11}
+	ldmia	sp!, {r7,r8,r9,r10,r11, pc}	@ return
+.size	vpaes_encrypt,.-vpaes_encrypt
+
+@
+@  Decryption stuff
+@
+.type	_vpaes_decrypt_consts,%object
+.align	4
+.Lk_dipt:@ decryption input transform
+.quad	0x0F505B040B545F00, 0x154A411E114E451A
+.quad	0x86E383E660056500, 0x12771772F491F194
+.Lk_dsbo:@ decryption sbox final output
+.quad	0x1387EA537EF94000, 0xC7AA6DB9D4943E2D
+.quad	0x12D7560F93441D00, 0xCA4B8159D8C58E9C
+.Lk_dsb9:@ decryption sbox output *9*u, *9*t
+.quad	0x851C03539A86D600, 0xCAD51F504F994CC9
+.quad	0xC03B1789ECD74900, 0x725E2C9EB2FBA565
+.Lk_dsbd:@ decryption sbox output *D*u, *D*t
+.quad	0x7D57CCDFE6B1A200, 0xF56E9B13882A4439
+.quad	0x3CE2FAF724C6CB00, 0x2931180D15DEEFD3
+.Lk_dsbb:@ decryption sbox output *B*u, *B*t
+.quad	0xD022649296B44200, 0x602646F6B0F2D404
+.quad	0xC19498A6CD596700, 0xF3FF0C3E3255AA6B
+.Lk_dsbe:@ decryption sbox output *E*u, *E*t
+.quad	0x46F2929626D4D000, 0x2242600464B4F6B0
+.quad	0x0C55A6CDFFAAC100, 0x9467F36B98593E32
+.size	_vpaes_decrypt_consts,.-_vpaes_decrypt_consts
+
+@@
+@@  Decryption core
+@@
+@@  Same API as encryption core, except it clobbers q12-q15 rather than using
+@@  the values from _vpaes_preheat. q9-q11 must still be set from
+@@  _vpaes_preheat.
+@@
+.type	_vpaes_decrypt_core,%function
+.align	4
+_vpaes_decrypt_core:
+	mov	r9, r2
+	ldr	r8, [r2,#240]		@ pull rounds
+
+	@ This function performs shuffles with various constants. The x86_64
+	@ version loads them on-demand into %xmm0-%xmm5. This does not work well
+	@ for ARMv7 because those registers are shuffle destinations. The ARMv8
+	@ version preloads those constants into registers, but ARMv7 has half
+	@ the registers to work with. Instead, we load them on-demand into
+	@ q12-q15, registers normally use for preloaded constants. This is fine
+	@ because decryption doesn't use those constants. The values are
+	@ constant, so this does not interfere with potential 2x optimizations.
+	adr	r7, .Lk_dipt
+
+	vld1.64	{q12,q13}, [r7]		@ vmovdqa	.Lk_dipt(%rip), %xmm2	# iptlo
+	lsl	r11, r8, #4		@ mov		%rax,	%r11;	shl	$4, %r11
+	eor	r11, r11, #0x30		@ xor		$0x30,	%r11
+	adr	r10, .Lk_sr
+	and	r11, r11, #0x30		@ and		$0x30,	%r11
+	add	r11, r11, r10
+	adr	r10, .Lk_mc_forward+48
+
+	vld1.64	{q4}, [r9]!		@ vmovdqu	(%r9),	%xmm4		# round0 key
+	vand	q1, q0, q9		@ vpand		%xmm9,	%xmm0,	%xmm1
+	vshr.u8	q0, q0, #4		@ vpsrlb	$4,	%xmm0,	%xmm0
+	vtbl.8	d4, {q12}, d2	@ vpshufb	%xmm1,	%xmm2,	%xmm2
+	vtbl.8	d5, {q12}, d3
+	vld1.64	{q5}, [r10]		@ vmovdqa	.Lk_mc_forward+48(%rip), %xmm5
+					@ vmovdqa	.Lk_dipt+16(%rip), %xmm1 # ipthi
+	vtbl.8	d0, {q13}, d0	@ vpshufb	%xmm0,	%xmm1,	%xmm0
+	vtbl.8	d1, {q13}, d1
+	veor	q2, q2, q4		@ vpxor		%xmm4,	%xmm2,	%xmm2
+	veor	q0, q0, q2		@ vpxor		%xmm2,	%xmm0,	%xmm0
+
+	@ .Ldec_entry ends with a bnz instruction which is normally paired with
+	@ subs in .Ldec_loop.
+	tst	r8, r8
+	b	.Ldec_entry
+
+.align	4
+.Ldec_loop:
+@
+@  Inverse mix columns
+@
+
+	@ We load .Lk_dsb* into q12-q15 on-demand. See the comment at the top of
+	@ the function.
+	adr	r10, .Lk_dsb9
+	vld1.64	{q12,q13}, [r10]!	@ vmovdqa	-0x20(%r10),%xmm4		# 4 : sb9u
+					@ vmovdqa	-0x10(%r10),%xmm1		# 0 : sb9t
+	@ Load sbd* ahead of time.
+	vld1.64	{q14,q15}, [r10]!	@ vmovdqa	0x00(%r10),%xmm4		# 4 : sbdu
+					@ vmovdqa	0x10(%r10),%xmm1		# 0 : sbdt
+	vtbl.8	d8, {q12}, d4	@ vpshufb	%xmm2,	%xmm4,	%xmm4		# 4 = sb9u
+	vtbl.8	d9, {q12}, d5
+	vtbl.8	d2, {q13}, d6	@ vpshufb	%xmm3,	%xmm1,	%xmm1		# 0 = sb9t
+	vtbl.8	d3, {q13}, d7
+	veor	q0, q4, q0		@ vpxor		%xmm4,	%xmm0,	%xmm0
+
+	veor	q0, q0, q1		@ vpxor		%xmm1,	%xmm0,	%xmm0		# 0 = ch
+
+	@ Load sbb* ahead of time.
+	vld1.64	{q12,q13}, [r10]!	@ vmovdqa	0x20(%r10),%xmm4		# 4 : sbbu
+					@ vmovdqa	0x30(%r10),%xmm1		# 0 : sbbt
+
+	vtbl.8	d8, {q14}, d4	@ vpshufb	%xmm2,	%xmm4,	%xmm4		# 4 = sbdu
+	vtbl.8	d9, {q14}, d5
+	@ Write to q1 instead of q0, so the table and destination registers do
+	@ not overlap.
+	vtbl.8	d2, {q0}, d10	@ vpshufb	%xmm5,	%xmm0,	%xmm0		# MC ch
+	vtbl.8	d3, {q0}, d11
+	@ Here we restore the original q0/q1 usage. This instruction is
+	@ reordered from the ARMv8 version so we do not clobber the vtbl.8
+	@ below.
+	veor	q0, q1, q4		@ vpxor		%xmm4,	%xmm0,	%xmm0		# 4 = ch
+	vtbl.8	d2, {q15}, d6	@ vpshufb	%xmm3,	%xmm1,	%xmm1		# 0 = sbdt
+	vtbl.8	d3, {q15}, d7
+					@ vmovdqa	0x20(%r10),	%xmm4		# 4 : sbbu
+	veor	q0, q0, q1		@ vpxor		%xmm1,	%xmm0,	%xmm0		# 0 = ch
+					@ vmovdqa	0x30(%r10),	%xmm1		# 0 : sbbt
+
+	@ Load sbd* ahead of time.
+	vld1.64	{q14,q15}, [r10]!	@ vmovdqa	0x40(%r10),%xmm4		# 4 : sbeu
+					@ vmovdqa	0x50(%r10),%xmm1		# 0 : sbet
+
+	vtbl.8	d8, {q12}, d4	@ vpshufb	%xmm2,	%xmm4,	%xmm4		# 4 = sbbu
+	vtbl.8	d9, {q12}, d5
+	@ Write to q1 instead of q0, so the table and destination registers do
+	@ not overlap.
+	vtbl.8	d2, {q0}, d10	@ vpshufb	%xmm5,	%xmm0,	%xmm0		# MC ch
+	vtbl.8	d3, {q0}, d11
+	@ Here we restore the original q0/q1 usage. This instruction is
+	@ reordered from the ARMv8 version so we do not clobber the vtbl.8
+	@ below.
+	veor	q0, q1, q4		@ vpxor		%xmm4,	%xmm0,	%xmm0		# 4 = ch
+	vtbl.8	d2, {q13}, d6	@ vpshufb	%xmm3,	%xmm1,	%xmm1		# 0 = sbbt
+	vtbl.8	d3, {q13}, d7
+	veor	q0, q0, q1		@ vpxor		%xmm1,	%xmm0,	%xmm0		# 0 = ch
+
+	vtbl.8	d8, {q14}, d4	@ vpshufb	%xmm2,	%xmm4,	%xmm4		# 4 = sbeu
+	vtbl.8	d9, {q14}, d5
+	@ Write to q1 instead of q0, so the table and destination registers do
+	@ not overlap.
+	vtbl.8	d2, {q0}, d10	@ vpshufb	%xmm5,	%xmm0,	%xmm0		# MC ch
+	vtbl.8	d3, {q0}, d11
+	@ Here we restore the original q0/q1 usage. This instruction is
+	@ reordered from the ARMv8 version so we do not clobber the vtbl.8
+	@ below.
+	veor	q0, q1, q4		@ vpxor		%xmm4,	%xmm0,	%xmm0		# 4 = ch
+	vtbl.8	d2, {q15}, d6	@ vpshufb	%xmm3,	%xmm1,	%xmm1		# 0 = sbet
+	vtbl.8	d3, {q15}, d7
+	vext.8	q5, q5, q5, #12		@ vpalignr 	$12,	%xmm5,	%xmm5,	%xmm5
+	veor	q0, q0, q1		@ vpxor		%xmm1,	%xmm0,	%xmm0		# 0 = ch
+	subs	r8, r8, #1		@ sub		$1,%rax			# nr--
+
+.Ldec_entry:
+	@ top of round
+	vand	q1, q0, q9		@ vpand		%xmm9,	%xmm0,	%xmm1	# 0 = k
+	vshr.u8	q0, q0, #4		@ vpsrlb	$4,	%xmm0,	%xmm0	# 1 = i
+	vtbl.8	d4, {q11}, d2	@ vpshufb	%xmm1,	%xmm11,	%xmm2	# 2 = a/k
+	vtbl.8	d5, {q11}, d3
+	veor	q1, q1, q0		@ vpxor		%xmm0,	%xmm1,	%xmm1	# 0 = j
+	vtbl.8	d6, {q10}, d0	@ vpshufb	%xmm0, 	%xmm10,	%xmm3	# 3 = 1/i
+	vtbl.8	d7, {q10}, d1
+	vtbl.8	d8, {q10}, d2	@ vpshufb	%xmm1,	%xmm10,	%xmm4	# 4 = 1/j
+	vtbl.8	d9, {q10}, d3
+	veor	q3, q3, q2		@ vpxor		%xmm2,	%xmm3,	%xmm3	# 3 = iak = 1/i + a/k
+	veor	q4, q4, q2		@ vpxor		%xmm2, 	%xmm4,	%xmm4	# 4 = jak = 1/j + a/k
+	vtbl.8	d4, {q10}, d6	@ vpshufb	%xmm3,	%xmm10,	%xmm2	# 2 = 1/iak
+	vtbl.8	d5, {q10}, d7
+	vtbl.8	d6, {q10}, d8	@ vpshufb	%xmm4,  %xmm10,	%xmm3	# 3 = 1/jak
+	vtbl.8	d7, {q10}, d9
+	veor	q2, q2, q1		@ vpxor		%xmm1,	%xmm2,	%xmm2	# 2 = io
+	veor	q3, q3, q0		@ vpxor		%xmm0,  %xmm3,	%xmm3	# 3 = jo
+	vld1.64	{q0}, [r9]!		@ vmovdqu	(%r9),	%xmm0
+	bne	.Ldec_loop
+
+	@ middle of last round
+
+	adr	r10, .Lk_dsbo
+
+	@ Write to q1 rather than q4 to avoid overlapping table and destination.
+	vld1.64	{q1}, [r10]!		@ vmovdqa	0x60(%r10),	%xmm4	# 3 : sbou
+	vtbl.8	d8, {q1}, d4	@ vpshufb	%xmm2,	%xmm4,	%xmm4	# 4 = sbou
+	vtbl.8	d9, {q1}, d5
+	@ Write to q2 rather than q1 to avoid overlapping table and destination.
+	vld1.64	{q2}, [r10]		@ vmovdqa	0x70(%r10),	%xmm1	# 0 : sbot
+	vtbl.8	d2, {q2}, d6	@ vpshufb	%xmm3,	%xmm1,	%xmm1	# 0 = sb1t
+	vtbl.8	d3, {q2}, d7
+	vld1.64	{q2}, [r11]		@ vmovdqa	-0x160(%r11),	%xmm2	# .Lk_sr-.Lk_dsbd=-0x160
+	veor	q4, q4, q0		@ vpxor		%xmm0,	%xmm4,	%xmm4	# 4 = sb1u + k
+	@ Write to q1 rather than q0 so the table and destination registers
+	@ below do not overlap.
+	veor	q1, q1, q4		@ vpxor		%xmm4,	%xmm1,	%xmm0	# 0 = A
+	vtbl.8	d0, {q1}, d4	@ vpshufb	%xmm2,	%xmm0,	%xmm0
+	vtbl.8	d1, {q1}, d5
+	bx	lr
+.size	_vpaes_decrypt_core,.-_vpaes_decrypt_core
+
+.globl	vpaes_decrypt
+.hidden	vpaes_decrypt
+.type	vpaes_decrypt,%function
+.align	4
+vpaes_decrypt:
+	@ _vpaes_decrypt_core uses r7-r11.
+	stmdb	sp!, {r7,r8,r9,r10,r11,lr}
+	@ _vpaes_decrypt_core uses q4-q5 (d8-d11), which are callee-saved.
+	vstmdb	sp!, {d8,d9,d10,d11}
+
+	vld1.64	{q0}, [r0]
+	bl	_vpaes_preheat
+	bl	_vpaes_decrypt_core
+	vst1.64	{q0}, [r1]
+
+	vldmia	sp!, {d8,d9,d10,d11}
+	ldmia	sp!, {r7,r8,r9,r10,r11, pc}	@ return
+.size	vpaes_decrypt,.-vpaes_decrypt
+@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
+@@                                                    @@
+@@                  AES key schedule                  @@
+@@                                                    @@
+@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
+
+@ This function diverges from both x86_64 and armv7 in which constants are
+@ pinned. x86_64 has a common preheat function for all operations. aarch64
+@ separates them because it has enough registers to pin nearly all constants.
+@ armv7 does not have enough registers, but needing explicit loads and stores
+@ also complicates using x86_64's register allocation directly.
+@
+@ We pin some constants for convenience and leave q14 and q15 free to load
+@ others on demand.
+
+@
+@  Key schedule constants
+@
+.type	_vpaes_key_consts,%object
+.align	4
+_vpaes_key_consts:
+.Lk_dksd:@ decryption key schedule: invskew x*D
+.quad	0xFEB91A5DA3E44700, 0x0740E3A45A1DBEF9
+.quad	0x41C277F4B5368300, 0x5FDC69EAAB289D1E
+.Lk_dksb:@ decryption key schedule: invskew x*B
+.quad	0x9A4FCA1F8550D500, 0x03D653861CC94C99
+.quad	0x115BEDA7B6FC4A00, 0xD993256F7E3482C8
+.Lk_dkse:@ decryption key schedule: invskew x*E + 0x63
+.quad	0xD5031CCA1FC9D600, 0x53859A4C994F5086
+.quad	0xA23196054FDC7BE8, 0xCD5EF96A20B31487
+.Lk_dks9:@ decryption key schedule: invskew x*9
+.quad	0xB6116FC87ED9A700, 0x4AED933482255BFC
+.quad	0x4576516227143300, 0x8BB89FACE9DAFDCE
+
+.Lk_rcon:@ rcon
+.quad	0x1F8391B9AF9DEEB6, 0x702A98084D7C7D81
+
+.Lk_opt:@ output transform
+.quad	0xFF9F4929D6B66000, 0xF7974121DEBE6808
+.quad	0x01EDBD5150BCEC00, 0xE10D5DB1B05C0CE0
+.Lk_deskew:@ deskew tables: inverts the sbox's "skew"
+.quad	0x07E4A34047A4E300, 0x1DFEB95A5DBEF91A
+.quad	0x5F36B5DC83EA6900, 0x2841C2ABF49D1E77
+.size	_vpaes_key_consts,.-_vpaes_key_consts
+
+.type	_vpaes_key_preheat,%function
+.align	4
+_vpaes_key_preheat:
+	adr	r11, .Lk_rcon
+	vmov.i8	q12, #0x5b			@ .Lk_s63
+	adr	r10, .Lk_inv			@ Must be aligned to 8 mod 16.
+	vmov.i8	q9, #0x0f			@ .Lk_s0F
+	vld1.64	{q10,q11}, [r10]		@ .Lk_inv
+	vld1.64	{q8}, [r11]			@ .Lk_rcon
+	bx	lr
+.size	_vpaes_key_preheat,.-_vpaes_key_preheat
+
+.type	_vpaes_schedule_core,%function
+.align	4
+_vpaes_schedule_core:
+	@ We only need to save lr, but ARM requires an 8-byte stack alignment,
+	@ so save an extra register.
+	stmdb	sp!, {r3,lr}
+
+	bl	_vpaes_key_preheat	@ load the tables
+
+	adr	r11, .Lk_ipt		@ Must be aligned to 8 mod 16.
+	vld1.64	{q0}, [r0]!		@ vmovdqu	(%rdi),	%xmm0		# load key (unaligned)
+
+	@ input transform
+	@ Use q4 here rather than q3 so .Lschedule_am_decrypting does not
+	@ overlap table and destination.
+	vmov	q4, q0			@ vmovdqa	%xmm0,	%xmm3
+	bl	_vpaes_schedule_transform
+	adr	r10, .Lk_sr		@ Must be aligned to 8 mod 16.
+	vmov	q7, q0			@ vmovdqa	%xmm0,	%xmm7
+
+	add	r8, r8, r10
+	tst	r3, r3
+	bne	.Lschedule_am_decrypting
+
+	@ encrypting, output zeroth round key after transform
+	vst1.64	{q0}, [r2]		@ vmovdqu	%xmm0,	(%rdx)
+	b	.Lschedule_go
+
+.Lschedule_am_decrypting:
+	@ decrypting, output zeroth round key after shiftrows
+	vld1.64	{q1}, [r8]		@ vmovdqa	(%r8,%r10),	%xmm1
+	vtbl.8	d6, {q4}, d2	@ vpshufb  	%xmm1,	%xmm3,	%xmm3
+	vtbl.8	d7, {q4}, d3
+	vst1.64	{q3}, [r2]		@ vmovdqu	%xmm3,	(%rdx)
+	eor	r8, r8, #0x30		@ xor	$0x30, %r8
+
+.Lschedule_go:
+	cmp	r1, #192		@ cmp	$192,	%esi
+	bhi	.Lschedule_256
+	beq	.Lschedule_192
+	@ 128: fall though
+
+@@
+@@  .schedule_128
+@@
+@@  128-bit specific part of key schedule.
+@@
+@@  This schedule is really simple, because all its parts
+@@  are accomplished by the subroutines.
+@@
+.Lschedule_128:
+	mov	r0, #10		@ mov	$10, %esi
+
+.Loop_schedule_128:
+	bl	_vpaes_schedule_round
+	subs	r0, r0, #1		@ dec	%esi
+	beq	.Lschedule_mangle_last
+	bl	_vpaes_schedule_mangle	@ write output
+	b	.Loop_schedule_128
+
+@@
+@@  .aes_schedule_192
+@@
+@@  192-bit specific part of key schedule.
+@@
+@@  The main body of this schedule is the same as the 128-bit
+@@  schedule, but with more smearing.  The long, high side is
+@@  stored in q7 as before, and the short, low side is in
+@@  the high bits of q6.
+@@
+@@  This schedule is somewhat nastier, however, because each
+@@  round produces 192 bits of key material, or 1.5 round keys.
+@@  Therefore, on each cycle we do 2 rounds and produce 3 round
+@@  keys.
+@@
+.align	4
+.Lschedule_192:
+	sub	r0, r0, #8
+	vld1.64	{q0}, [r0]			@ vmovdqu	8(%rdi),%xmm0		# load key part 2 (very unaligned)
+	bl	_vpaes_schedule_transform	@ input transform
+	vmov	q6, q0				@ vmovdqa	%xmm0,	%xmm6		# save short part
+	vmov.i8	d12, #0			@ vpxor	%xmm4,	%xmm4, %xmm4	# clear 4
+						@ vmovhlps	%xmm4,	%xmm6,	%xmm6		# clobber low side with zeros
+	mov	r0, #4			@ mov	$4,	%esi
+
+.Loop_schedule_192:
+	bl	_vpaes_schedule_round
+	vext.8	q0, q6, q0, #8			@ vpalignr	$8,%xmm6,%xmm0,%xmm0
+	bl	_vpaes_schedule_mangle		@ save key n
+	bl	_vpaes_schedule_192_smear
+	bl	_vpaes_schedule_mangle		@ save key n+1
+	bl	_vpaes_schedule_round
+	subs	r0, r0, #1			@ dec	%esi
+	beq	.Lschedule_mangle_last
+	bl	_vpaes_schedule_mangle		@ save key n+2
+	bl	_vpaes_schedule_192_smear
+	b	.Loop_schedule_192
+
+@@
+@@  .aes_schedule_256
+@@
+@@  256-bit specific part of key schedule.
+@@
+@@  The structure here is very similar to the 128-bit
+@@  schedule, but with an additional "low side" in
+@@  q6.  The low side's rounds are the same as the
+@@  high side's, except no rcon and no rotation.
+@@
+.align	4
+.Lschedule_256:
+	vld1.64	{q0}, [r0]			@ vmovdqu	16(%rdi),%xmm0		# load key part 2 (unaligned)
+	bl	_vpaes_schedule_transform	@ input transform
+	mov	r0, #7			@ mov	$7, %esi
+
+.Loop_schedule_256:
+	bl	_vpaes_schedule_mangle		@ output low result
+	vmov	q6, q0				@ vmovdqa	%xmm0,	%xmm6		# save cur_lo in xmm6
+
+	@ high round
+	bl	_vpaes_schedule_round
+	subs	r0, r0, #1			@ dec	%esi
+	beq	.Lschedule_mangle_last
+	bl	_vpaes_schedule_mangle
+
+	@ low round. swap xmm7 and xmm6
+	vdup.32	q0, d1[1]		@ vpshufd	$0xFF,	%xmm0,	%xmm0
+	vmov.i8	q4, #0
+	vmov	q5, q7			@ vmovdqa	%xmm7,	%xmm5
+	vmov	q7, q6			@ vmovdqa	%xmm6,	%xmm7
+	bl	_vpaes_schedule_low_round
+	vmov	q7, q5			@ vmovdqa	%xmm5,	%xmm7
+
+	b	.Loop_schedule_256
+
+@@
+@@  .aes_schedule_mangle_last
+@@
+@@  Mangler for last round of key schedule
+@@  Mangles q0
+@@    when encrypting, outputs out(q0) ^ 63
+@@    when decrypting, outputs unskew(q0)
+@@
+@@  Always called right before return... jumps to cleanup and exits
+@@
+.align	4
+.Lschedule_mangle_last:
+	@ schedule last round key from xmm0
+	adr	r11, .Lk_deskew			@ lea	.Lk_deskew(%rip),%r11	# prepare to deskew
+	tst	r3, r3
+	bne	.Lschedule_mangle_last_dec
+
+	@ encrypting
+	vld1.64	{q1}, [r8]		@ vmovdqa	(%r8,%r10),%xmm1
+	adr	r11, .Lk_opt		@ lea		.Lk_opt(%rip),	%r11		# prepare to output transform
+	add	r2, r2, #32		@ add		$32,	%rdx
+	vmov	q2, q0
+	vtbl.8	d0, {q2}, d2	@ vpshufb	%xmm1,	%xmm0,	%xmm0		# output permute
+	vtbl.8	d1, {q2}, d3
+
+.Lschedule_mangle_last_dec:
+	sub	r2, r2, #16			@ add	$-16,	%rdx
+	veor	q0, q0, q12			@ vpxor	.Lk_s63(%rip),	%xmm0,	%xmm0
+	bl	_vpaes_schedule_transform	@ output transform
+	vst1.64	{q0}, [r2]			@ vmovdqu	%xmm0,	(%rdx)		# save last key
+
+	@ cleanup
+	veor	q0, q0, q0		@ vpxor	%xmm0,	%xmm0,	%xmm0
+	veor	q1, q1, q1		@ vpxor	%xmm1,	%xmm1,	%xmm1
+	veor	q2, q2, q2		@ vpxor	%xmm2,	%xmm2,	%xmm2
+	veor	q3, q3, q3		@ vpxor	%xmm3,	%xmm3,	%xmm3
+	veor	q4, q4, q4		@ vpxor	%xmm4,	%xmm4,	%xmm4
+	veor	q5, q5, q5		@ vpxor	%xmm5,	%xmm5,	%xmm5
+	veor	q6, q6, q6		@ vpxor	%xmm6,	%xmm6,	%xmm6
+	veor	q7, q7, q7		@ vpxor	%xmm7,	%xmm7,	%xmm7
+	ldmia	sp!, {r3,pc}		@ return
+.size	_vpaes_schedule_core,.-_vpaes_schedule_core
+
+@@
+@@  .aes_schedule_192_smear
+@@
+@@  Smear the short, low side in the 192-bit key schedule.
+@@
+@@  Inputs:
+@@    q7: high side, b  a  x  y
+@@    q6:  low side, d  c  0  0
+@@
+@@  Outputs:
+@@    q6: b+c+d  b+c  0  0
+@@    q0: b+c+d  b+c  b  a
+@@
+.type	_vpaes_schedule_192_smear,%function
+.align	4
+_vpaes_schedule_192_smear:
+	vmov.i8	q1, #0
+	vdup.32	q0, d15[1]
+	vshl.i64	q1, q6, #32		@ vpshufd	$0x80,	%xmm6,	%xmm1	# d c 0 0 -> c 0 0 0
+	vmov	d0, d15		@ vpshufd	$0xFE,	%xmm7,	%xmm0	# b a _ _ -> b b b a
+	veor	q6, q6, q1		@ vpxor	%xmm1,	%xmm6,	%xmm6	# -> c+d c 0 0
+	veor	q1, q1, q1		@ vpxor	%xmm1,	%xmm1,	%xmm1
+	veor	q6, q6, q0		@ vpxor	%xmm0,	%xmm6,	%xmm6	# -> b+c+d b+c b a
+	vmov	q0, q6			@ vmovdqa	%xmm6,	%xmm0
+	vmov	d12, d2		@ vmovhlps	%xmm1,	%xmm6,	%xmm6	# clobber low side with zeros
+	bx	lr
+.size	_vpaes_schedule_192_smear,.-_vpaes_schedule_192_smear
+
+@@
+@@  .aes_schedule_round
+@@
+@@  Runs one main round of the key schedule on q0, q7
+@@
+@@  Specifically, runs subbytes on the high dword of q0
+@@  then rotates it by one byte and xors into the low dword of
+@@  q7.
+@@
+@@  Adds rcon from low byte of q8, then rotates q8 for
+@@  next rcon.
+@@
+@@  Smears the dwords of q7 by xoring the low into the
+@@  second low, result into third, result into highest.
+@@
+@@  Returns results in q7 = q0.
+@@  Clobbers q1-q4, r11.
+@@
+.type	_vpaes_schedule_round,%function
+.align	4
+_vpaes_schedule_round:
+	@ extract rcon from xmm8
+	vmov.i8	q4, #0				@ vpxor		%xmm4,	%xmm4,	%xmm4
+	vext.8	q1, q8, q4, #15		@ vpalignr	$15,	%xmm8,	%xmm4,	%xmm1
+	vext.8	q8, q8, q8, #15	@ vpalignr	$15,	%xmm8,	%xmm8,	%xmm8
+	veor	q7, q7, q1			@ vpxor		%xmm1,	%xmm7,	%xmm7
+
+	@ rotate
+	vdup.32	q0, d1[1]			@ vpshufd	$0xFF,	%xmm0,	%xmm0
+	vext.8	q0, q0, q0, #1			@ vpalignr	$1,	%xmm0,	%xmm0,	%xmm0
+
+	@ fall through...
+
+	@ low round: same as high round, but no rotation and no rcon.
+_vpaes_schedule_low_round:
+	@ The x86_64 version pins .Lk_sb1 in %xmm13 and .Lk_sb1+16 in %xmm12.
+	@ We pin other values in _vpaes_key_preheat, so load them now.
+	adr	r11, .Lk_sb1
+	vld1.64	{q14,q15}, [r11]
+
+	@ smear xmm7
+	vext.8	q1, q4, q7, #12			@ vpslldq	$4,	%xmm7,	%xmm1
+	veor	q7, q7, q1			@ vpxor	%xmm1,	%xmm7,	%xmm7
+	vext.8	q4, q4, q7, #8			@ vpslldq	$8,	%xmm7,	%xmm4
+
+	@ subbytes
+	vand	q1, q0, q9			@ vpand		%xmm9,	%xmm0,	%xmm1		# 0 = k
+	vshr.u8	q0, q0, #4			@ vpsrlb	$4,	%xmm0,	%xmm0		# 1 = i
+	veor	q7, q7, q4			@ vpxor		%xmm4,	%xmm7,	%xmm7
+	vtbl.8	d4, {q11}, d2		@ vpshufb	%xmm1,	%xmm11,	%xmm2		# 2 = a/k
+	vtbl.8	d5, {q11}, d3
+	veor	q1, q1, q0			@ vpxor		%xmm0,	%xmm1,	%xmm1		# 0 = j
+	vtbl.8	d6, {q10}, d0		@ vpshufb	%xmm0, 	%xmm10,	%xmm3		# 3 = 1/i
+	vtbl.8	d7, {q10}, d1
+	veor	q3, q3, q2			@ vpxor		%xmm2,	%xmm3,	%xmm3		# 3 = iak = 1/i + a/k
+	vtbl.8	d8, {q10}, d2		@ vpshufb	%xmm1,	%xmm10,	%xmm4		# 4 = 1/j
+	vtbl.8	d9, {q10}, d3
+	veor	q7, q7, q12			@ vpxor		.Lk_s63(%rip),	%xmm7,	%xmm7
+	vtbl.8	d6, {q10}, d6		@ vpshufb	%xmm3,	%xmm10,	%xmm3		# 2 = 1/iak
+	vtbl.8	d7, {q10}, d7
+	veor	q4, q4, q2			@ vpxor		%xmm2,	%xmm4,	%xmm4		# 4 = jak = 1/j + a/k
+	vtbl.8	d4, {q10}, d8		@ vpshufb	%xmm4,	%xmm10,	%xmm2		# 3 = 1/jak
+	vtbl.8	d5, {q10}, d9
+	veor	q3, q3, q1			@ vpxor		%xmm1,	%xmm3,	%xmm3		# 2 = io
+	veor	q2, q2, q0			@ vpxor		%xmm0,	%xmm2,	%xmm2		# 3 = jo
+	vtbl.8	d8, {q15}, d6		@ vpshufb	%xmm3,	%xmm13,	%xmm4		# 4 = sbou
+	vtbl.8	d9, {q15}, d7
+	vtbl.8	d2, {q14}, d4		@ vpshufb	%xmm2,	%xmm12,	%xmm1		# 0 = sb1t
+	vtbl.8	d3, {q14}, d5
+	veor	q1, q1, q4			@ vpxor		%xmm4,	%xmm1,	%xmm1		# 0 = sbox output
+
+	@ add in smeared stuff
+	veor	q0, q1, q7			@ vpxor	%xmm7,	%xmm1,	%xmm0
+	veor	q7, q1, q7			@ vmovdqa	%xmm0,	%xmm7
+	bx	lr
+.size	_vpaes_schedule_round,.-_vpaes_schedule_round
+
+@@
+@@  .aes_schedule_transform
+@@
+@@  Linear-transform q0 according to tables at [r11]
+@@
+@@  Requires that q9 = 0x0F0F... as in preheat
+@@  Output in q0
+@@  Clobbers q1, q2, q14, q15
+@@
+.type	_vpaes_schedule_transform,%function
+.align	4
+_vpaes_schedule_transform:
+	vld1.64	{q14,q15}, [r11]	@ vmovdqa	(%r11),	%xmm2 	# lo
+					@ vmovdqa	16(%r11),	%xmm1 # hi
+	vand	q1, q0, q9		@ vpand	%xmm9,	%xmm0,	%xmm1
+	vshr.u8	q0, q0, #4		@ vpsrlb	$4,	%xmm0,	%xmm0
+	vtbl.8	d4, {q14}, d2	@ vpshufb	%xmm1,	%xmm2,	%xmm2
+	vtbl.8	d5, {q14}, d3
+	vtbl.8	d0, {q15}, d0	@ vpshufb	%xmm0,	%xmm1,	%xmm0
+	vtbl.8	d1, {q15}, d1
+	veor	q0, q0, q2		@ vpxor	%xmm2,	%xmm0,	%xmm0
+	bx	lr
+.size	_vpaes_schedule_transform,.-_vpaes_schedule_transform
+
+@@
+@@  .aes_schedule_mangle
+@@
+@@  Mangles q0 from (basis-transformed) standard version
+@@  to our version.
+@@
+@@  On encrypt,
+@@    xor with 0x63
+@@    multiply by circulant 0,1,1,1
+@@    apply shiftrows transform
+@@
+@@  On decrypt,
+@@    xor with 0x63
+@@    multiply by "inverse mixcolumns" circulant E,B,D,9
+@@    deskew
+@@    apply shiftrows transform
+@@
+@@
+@@  Writes out to [r2], and increments or decrements it
+@@  Keeps track of round number mod 4 in r8
+@@  Preserves q0
+@@  Clobbers q1-q5
+@@
+.type	_vpaes_schedule_mangle,%function
+.align	4
+_vpaes_schedule_mangle:
+	tst	r3, r3
+	vmov	q4, q0			@ vmovdqa	%xmm0,	%xmm4	# save xmm0 for later
+	adr	r11, .Lk_mc_forward	@ Must be aligned to 8 mod 16.
+	vld1.64	{q5}, [r11]		@ vmovdqa	.Lk_mc_forward(%rip),%xmm5
+	bne	.Lschedule_mangle_dec
+
+	@ encrypting
+	@ Write to q2 so we do not overlap table and destination below.
+	veor	q2, q0, q12		@ vpxor		.Lk_s63(%rip),	%xmm0,	%xmm4
+	add	r2, r2, #16		@ add		$16,	%rdx
+	vtbl.8	d8, {q2}, d10	@ vpshufb	%xmm5,	%xmm4,	%xmm4
+	vtbl.8	d9, {q2}, d11
+	vtbl.8	d2, {q4}, d10	@ vpshufb	%xmm5,	%xmm4,	%xmm1
+	vtbl.8	d3, {q4}, d11
+	vtbl.8	d6, {q1}, d10	@ vpshufb	%xmm5,	%xmm1,	%xmm3
+	vtbl.8	d7, {q1}, d11
+	veor	q4, q4, q1		@ vpxor		%xmm1,	%xmm4,	%xmm4
+	vld1.64	{q1}, [r8]		@ vmovdqa	(%r8,%r10),	%xmm1
+	veor	q3, q3, q4		@ vpxor		%xmm4,	%xmm3,	%xmm3
+
+	b	.Lschedule_mangle_both
+.align	4
+.Lschedule_mangle_dec:
+	@ inverse mix columns
+	adr	r11, .Lk_dksd 		@ lea		.Lk_dksd(%rip),%r11
+	vshr.u8	q1, q4, #4		@ vpsrlb	$4,	%xmm4,	%xmm1	# 1 = hi
+	vand	q4, q4, q9		@ vpand		%xmm9,	%xmm4,	%xmm4	# 4 = lo
+
+	vld1.64	{q14,q15}, [r11]! 	@ vmovdqa	0x00(%r11),	%xmm2
+					@ vmovdqa	0x10(%r11),	%xmm3
+	vtbl.8	d4, {q14}, d8	@ vpshufb	%xmm4,	%xmm2,	%xmm2
+	vtbl.8	d5, {q14}, d9
+	vtbl.8	d6, {q15}, d2	@ vpshufb	%xmm1,	%xmm3,	%xmm3
+	vtbl.8	d7, {q15}, d3
+	@ Load .Lk_dksb ahead of time.
+	vld1.64	{q14,q15}, [r11]! 	@ vmovdqa	0x20(%r11),	%xmm2
+					@ vmovdqa	0x30(%r11),	%xmm3
+	@ Write to q13 so we do not overlap table and destination.
+	veor	q13, q3, q2		@ vpxor		%xmm2,	%xmm3,	%xmm3
+	vtbl.8	d6, {q13}, d10	@ vpshufb	%xmm5,	%xmm3,	%xmm3
+	vtbl.8	d7, {q13}, d11
+
+	vtbl.8	d4, {q14}, d8	@ vpshufb	%xmm4,	%xmm2,	%xmm2
+	vtbl.8	d5, {q14}, d9
+	veor	q2, q2, q3		@ vpxor		%xmm3,	%xmm2,	%xmm2
+	vtbl.8	d6, {q15}, d2	@ vpshufb	%xmm1,	%xmm3,	%xmm3
+	vtbl.8	d7, {q15}, d3
+	@ Load .Lk_dkse ahead of time.
+	vld1.64	{q14,q15}, [r11]! 	@ vmovdqa	0x40(%r11),	%xmm2
+					@ vmovdqa	0x50(%r11),	%xmm3
+	@ Write to q13 so we do not overlap table and destination.
+	veor	q13, q3, q2		@ vpxor		%xmm2,	%xmm3,	%xmm3
+	vtbl.8	d6, {q13}, d10	@ vpshufb	%xmm5,	%xmm3,	%xmm3
+	vtbl.8	d7, {q13}, d11
+
+	vtbl.8	d4, {q14}, d8	@ vpshufb	%xmm4,	%xmm2,	%xmm2
+	vtbl.8	d5, {q14}, d9
+	veor	q2, q2, q3		@ vpxor		%xmm3,	%xmm2,	%xmm2
+	vtbl.8	d6, {q15}, d2	@ vpshufb	%xmm1,	%xmm3,	%xmm3
+	vtbl.8	d7, {q15}, d3
+	@ Load .Lk_dkse ahead of time.
+	vld1.64	{q14,q15}, [r11]! 	@ vmovdqa	0x60(%r11),	%xmm2
+					@ vmovdqa	0x70(%r11),	%xmm4
+	@ Write to q13 so we do not overlap table and destination.
+	veor	q13, q3, q2		@ vpxor		%xmm2,	%xmm3,	%xmm3
+
+	vtbl.8	d4, {q14}, d8	@ vpshufb	%xmm4,	%xmm2,	%xmm2
+	vtbl.8	d5, {q14}, d9
+	vtbl.8	d6, {q13}, d10	@ vpshufb	%xmm5,	%xmm3,	%xmm3
+	vtbl.8	d7, {q13}, d11
+	vtbl.8	d8, {q15}, d2	@ vpshufb	%xmm1,	%xmm4,	%xmm4
+	vtbl.8	d9, {q15}, d3
+	vld1.64	{q1}, [r8]		@ vmovdqa	(%r8,%r10),	%xmm1
+	veor	q2, q2, q3		@ vpxor	%xmm3,	%xmm2,	%xmm2
+	veor	q3, q4, q2		@ vpxor	%xmm2,	%xmm4,	%xmm3
+
+	sub	r2, r2, #16		@ add	$-16,	%rdx
+
+.Lschedule_mangle_both:
+	@ Write to q2 so table and destination do not overlap.
+	vtbl.8	d4, {q3}, d2	@ vpshufb	%xmm1,	%xmm3,	%xmm3
+	vtbl.8	d5, {q3}, d3
+	add	r8, r8, #64-16		@ add	$-16,	%r8
+	and	r8, r8, #~(1<<6)	@ and	$0x30,	%r8
+	vst1.64	{q2}, [r2]		@ vmovdqu	%xmm3,	(%rdx)
+	bx	lr
+.size	_vpaes_schedule_mangle,.-_vpaes_schedule_mangle
+
+.globl	vpaes_set_encrypt_key
+.hidden	vpaes_set_encrypt_key
+.type	vpaes_set_encrypt_key,%function
+.align	4
+vpaes_set_encrypt_key:
+	stmdb	sp!, {r7,r8,r9,r10,r11, lr}
+	vstmdb	sp!, {d8,d9,d10,d11,d12,d13,d14,d15}
+
+	lsr	r9, r1, #5		@ shr	$5,%eax
+	add	r9, r9, #5		@ $5,%eax
+	str	r9, [r2,#240]		@ mov	%eax,240(%rdx)	# AES_KEY->rounds = nbits/32+5;
+
+	mov	r3, #0		@ mov	$0,%ecx
+	mov	r8, #0x30		@ mov	$0x30,%r8d
+	bl	_vpaes_schedule_core
+	eor	r0, r0, r0
+
+	vldmia	sp!, {d8,d9,d10,d11,d12,d13,d14,d15}
+	ldmia	sp!, {r7,r8,r9,r10,r11, pc}	@ return
+.size	vpaes_set_encrypt_key,.-vpaes_set_encrypt_key
+
+.globl	vpaes_set_decrypt_key
+.hidden	vpaes_set_decrypt_key
+.type	vpaes_set_decrypt_key,%function
+.align	4
+vpaes_set_decrypt_key:
+	stmdb	sp!, {r7,r8,r9,r10,r11, lr}
+	vstmdb	sp!, {d8,d9,d10,d11,d12,d13,d14,d15}
+
+	lsr	r9, r1, #5		@ shr	$5,%eax
+	add	r9, r9, #5		@ $5,%eax
+	str	r9, [r2,#240]		@ mov	%eax,240(%rdx)	# AES_KEY->rounds = nbits/32+5;
+	lsl	r9, r9, #4		@ shl	$4,%eax
+	add	r2, r2, #16		@ lea	16(%rdx,%rax),%rdx
+	add	r2, r2, r9
+
+	mov	r3, #1		@ mov	$1,%ecx
+	lsr	r8, r1, #1		@ shr	$1,%r8d
+	and	r8, r8, #32		@ and	$32,%r8d
+	eor	r8, r8, #32		@ xor	$32,%r8d	# nbits==192?0:32
+	bl	_vpaes_schedule_core
+
+	vldmia	sp!, {d8,d9,d10,d11,d12,d13,d14,d15}
+	ldmia	sp!, {r7,r8,r9,r10,r11, pc}	@ return
+.size	vpaes_set_decrypt_key,.-vpaes_set_decrypt_key
+
+@ Additional constants for converting to bsaes.
+.type	_vpaes_convert_consts,%object
+.align	4
+_vpaes_convert_consts:
+@ .Lk_opt_then_skew applies skew(opt(x)) XOR 0x63, where skew is the linear
+@ transform in the AES S-box. 0x63 is incorporated into the low half of the
+@ table. This was computed with the following script:
+@
+@   def u64s_to_u128(x, y):
+@       return x | (y << 64)
+@   def u128_to_u64s(w):
+@       return w & ((1<<64)-1), w >> 64
+@   def get_byte(w, i):
+@       return (w >> (i*8)) & 0xff
+@   def apply_table(table, b):
+@       lo = b & 0xf
+@       hi = b >> 4
+@       return get_byte(table[0], lo) ^ get_byte(table[1], hi)
+@   def opt(b):
+@       table = [
+@           u64s_to_u128(0xFF9F4929D6B66000, 0xF7974121DEBE6808),
+@           u64s_to_u128(0x01EDBD5150BCEC00, 0xE10D5DB1B05C0CE0),
+@       ]
+@       return apply_table(table, b)
+@   def rot_byte(b, n):
+@       return 0xff & ((b << n) | (b >> (8-n)))
+@   def skew(x):
+@       return (x ^ rot_byte(x, 1) ^ rot_byte(x, 2) ^ rot_byte(x, 3) ^
+@               rot_byte(x, 4))
+@   table = [0, 0]
+@   for i in range(16):
+@       table[0] |= (skew(opt(i)) ^ 0x63) << (i*8)
+@       table[1] |= skew(opt(i<<4)) << (i*8)
+@   print("	.quad	0x%016x, 0x%016x" % u128_to_u64s(table[0]))
+@   print("	.quad	0x%016x, 0x%016x" % u128_to_u64s(table[1]))
+.Lk_opt_then_skew:
+.quad	0x9cb8436798bc4763, 0x6440bb9f6044bf9b
+.quad	0x1f30062936192f00, 0xb49bad829db284ab
+
+@ .Lk_decrypt_transform is a permutation which performs an 8-bit left-rotation
+@ followed by a byte-swap on each 32-bit word of a vector. E.g., 0x11223344
+@ becomes 0x22334411 and then 0x11443322.
+.Lk_decrypt_transform:
+.quad	0x0704050603000102, 0x0f0c0d0e0b08090a
+.size	_vpaes_convert_consts,.-_vpaes_convert_consts
+
+@ void vpaes_encrypt_key_to_bsaes(AES_KEY *bsaes, const AES_KEY *vpaes);
+.globl	vpaes_encrypt_key_to_bsaes
+.hidden	vpaes_encrypt_key_to_bsaes
+.type	vpaes_encrypt_key_to_bsaes,%function
+.align	4
+vpaes_encrypt_key_to_bsaes:
+	stmdb	sp!, {r11, lr}
+
+	@ See _vpaes_schedule_core for the key schedule logic. In particular,
+	@ _vpaes_schedule_transform(.Lk_ipt) (section 2.2 of the paper),
+	@ _vpaes_schedule_mangle (section 4.3), and .Lschedule_mangle_last
+	@ contain the transformations not in the bsaes representation. This
+	@ function inverts those transforms.
+	@
+	@ Note also that bsaes-armv7.pl expects aes-armv4.pl's key
+	@ representation, which does not match the other aes_nohw_*
+	@ implementations. The ARM aes_nohw_* stores each 32-bit word
+	@ byteswapped, as a convenience for (unsupported) big-endian ARM, at the
+	@ cost of extra REV and VREV32 operations in little-endian ARM.
+
+	vmov.i8	q9, #0x0f		@ Required by _vpaes_schedule_transform
+	adr	r2, .Lk_mc_forward	@ Must be aligned to 8 mod 16.
+	add	r3, r2, 0x90		@ .Lk_sr+0x10-.Lk_mc_forward = 0x90 (Apple's toolchain doesn't support the expression)
+
+	vld1.64	{q12}, [r2]
+	vmov.i8	q10, #0x5b		@ .Lk_s63 from vpaes-x86_64
+	adr	r11, .Lk_opt		@ Must be aligned to 8 mod 16.
+	vmov.i8	q11, #0x63		@ .LK_s63 without .Lk_ipt applied
+
+	@ vpaes stores one fewer round count than bsaes, but the number of keys
+	@ is the same.
+	ldr	r2, [r1,#240]
+	add	r2, r2, #1
+	str	r2, [r0,#240]
+
+	@ The first key is transformed with _vpaes_schedule_transform(.Lk_ipt).
+	@ Invert this with .Lk_opt.
+	vld1.64	{q0}, [r1]!
+	bl	_vpaes_schedule_transform
+	vrev32.8	q0, q0
+	vst1.64	{q0}, [r0]!
+
+	@ The middle keys have _vpaes_schedule_transform(.Lk_ipt) applied,
+	@ followed by _vpaes_schedule_mangle. _vpaes_schedule_mangle XORs 0x63,
+	@ multiplies by the circulant 0,1,1,1, then applies ShiftRows.
+.Loop_enc_key_to_bsaes:
+	vld1.64	{q0}, [r1]!
+
+	@ Invert the ShiftRows step (see .Lschedule_mangle_both). Note we cycle
+	@ r3 in the opposite direction and start at .Lk_sr+0x10 instead of 0x30.
+	@ We use r3 rather than r8 to avoid a callee-saved register.
+	vld1.64	{q1}, [r3]
+	vtbl.8	d4, {q0}, d2
+	vtbl.8	d5, {q0}, d3
+	add	r3, r3, #16
+	and	r3, r3, #~(1<<6)
+	vmov	q0, q2
+
+	@ Handle the last key differently.
+	subs	r2, r2, #1
+	beq	.Loop_enc_key_to_bsaes_last
+
+	@ Multiply by the circulant. This is its own inverse.
+	vtbl.8	d2, {q0}, d24
+	vtbl.8	d3, {q0}, d25
+	vmov	q0, q1
+	vtbl.8	d4, {q1}, d24
+	vtbl.8	d5, {q1}, d25
+	veor	q0, q0, q2
+	vtbl.8	d2, {q2}, d24
+	vtbl.8	d3, {q2}, d25
+	veor	q0, q0, q1
+
+	@ XOR and finish.
+	veor	q0, q0, q10
+	bl	_vpaes_schedule_transform
+	vrev32.8	q0, q0
+	vst1.64	{q0}, [r0]!
+	b	.Loop_enc_key_to_bsaes
+
+.Loop_enc_key_to_bsaes_last:
+	@ The final key does not have a basis transform (note
+	@ .Lschedule_mangle_last inverts the original transform). It only XORs
+	@ 0x63 and applies ShiftRows. The latter was already inverted in the
+	@ loop. Note that, because we act on the original representation, we use
+	@ q11, not q10.
+	veor	q0, q0, q11
+	vrev32.8	q0, q0
+	vst1.64	{q0}, [r0]
+
+	@ Wipe registers which contained key material.
+	veor	q0, q0, q0
+	veor	q1, q1, q1
+	veor	q2, q2, q2
+
+	ldmia	sp!, {r11, pc}	@ return
+.size	vpaes_encrypt_key_to_bsaes,.-vpaes_encrypt_key_to_bsaes
+
+@ void vpaes_decrypt_key_to_bsaes(AES_KEY *vpaes, const AES_KEY *bsaes);
+.globl	vpaes_decrypt_key_to_bsaes
+.hidden	vpaes_decrypt_key_to_bsaes
+.type	vpaes_decrypt_key_to_bsaes,%function
+.align	4
+vpaes_decrypt_key_to_bsaes:
+	stmdb	sp!, {r11, lr}
+
+	@ See _vpaes_schedule_core for the key schedule logic. Note vpaes
+	@ computes the decryption key schedule in reverse. Additionally,
+	@ aes-x86_64.pl shares some transformations, so we must only partially
+	@ invert vpaes's transformations. In general, vpaes computes in a
+	@ different basis (.Lk_ipt and .Lk_opt) and applies the inverses of
+	@ MixColumns, ShiftRows, and the affine part of the AES S-box (which is
+	@ split into a linear skew and XOR of 0x63). We undo all but MixColumns.
+	@
+	@ Note also that bsaes-armv7.pl expects aes-armv4.pl's key
+	@ representation, which does not match the other aes_nohw_*
+	@ implementations. The ARM aes_nohw_* stores each 32-bit word
+	@ byteswapped, as a convenience for (unsupported) big-endian ARM, at the
+	@ cost of extra REV and VREV32 operations in little-endian ARM.
+
+	adr	r2, .Lk_decrypt_transform
+	adr	r3, .Lk_sr+0x30
+	adr	r11, .Lk_opt_then_skew	@ Input to _vpaes_schedule_transform.
+	vld1.64	{q12}, [r2]	@ Reuse q12 from encryption.
+	vmov.i8	q9, #0x0f		@ Required by _vpaes_schedule_transform
+
+	@ vpaes stores one fewer round count than bsaes, but the number of keys
+	@ is the same.
+	ldr	r2, [r1,#240]
+	add	r2, r2, #1
+	str	r2, [r0,#240]
+
+	@ Undo the basis change and reapply the S-box affine transform. See
+	@ .Lschedule_mangle_last.
+	vld1.64	{q0}, [r1]!
+	bl	_vpaes_schedule_transform
+	vrev32.8	q0, q0
+	vst1.64	{q0}, [r0]!
+
+	@ See _vpaes_schedule_mangle for the transform on the middle keys. Note
+	@ it simultaneously inverts MixColumns and the S-box affine transform.
+	@ See .Lk_dksd through .Lk_dks9.
+.Loop_dec_key_to_bsaes:
+	vld1.64	{q0}, [r1]!
+
+	@ Invert the ShiftRows step (see .Lschedule_mangle_both). Note going
+	@ forwards cancels inverting for which direction we cycle r3. We use r3
+	@ rather than r8 to avoid a callee-saved register.
+	vld1.64	{q1}, [r3]
+	vtbl.8	d4, {q0}, d2
+	vtbl.8	d5, {q0}, d3
+	add	r3, r3, #64-16
+	and	r3, r3, #~(1<<6)
+	vmov	q0, q2
+
+	@ Handle the last key differently.
+	subs	r2, r2, #1
+	beq	.Loop_dec_key_to_bsaes_last
+
+	@ Undo the basis change and reapply the S-box affine transform.
+	bl	_vpaes_schedule_transform
+
+	@ Rotate each word by 8 bytes (cycle the rows) and then byte-swap. We
+	@ combine the two operations in .Lk_decrypt_transform.
+	@
+	@ TODO(davidben): Where does the rotation come from?
+	vtbl.8	d2, {q0}, d24
+	vtbl.8	d3, {q0}, d25
+
+	vst1.64	{q1}, [r0]!
+	b	.Loop_dec_key_to_bsaes
+
+.Loop_dec_key_to_bsaes_last:
+	@ The final key only inverts ShiftRows (already done in the loop). See
+	@ .Lschedule_am_decrypting. Its basis is not transformed.
+	vrev32.8	q0, q0
+	vst1.64	{q0}, [r0]!
+
+	@ Wipe registers which contained key material.
+	veor	q0, q0, q0
+	veor	q1, q1, q1
+	veor	q2, q2, q2
+
+	ldmia	sp!, {r11, pc}	@ return
+.size	vpaes_decrypt_key_to_bsaes,.-vpaes_decrypt_key_to_bsaes
+.globl	vpaes_ctr32_encrypt_blocks
+.hidden	vpaes_ctr32_encrypt_blocks
+.type	vpaes_ctr32_encrypt_blocks,%function
+.align	4
+vpaes_ctr32_encrypt_blocks:
+	mov	ip, sp
+	stmdb	sp!, {r7,r8,r9,r10,r11, lr}
+	@ This function uses q4-q7 (d8-d15), which are callee-saved.
+	vstmdb	sp!, {d8,d9,d10,d11,d12,d13,d14,d15}
+
+	cmp	r2, #0
+	@ r8 is passed on the stack.
+	ldr	r8, [ip]
+	beq	.Lctr32_done
+
+	@ _vpaes_encrypt_core expects the key in r2, so swap r2 and r3.
+	mov	r9, r3
+	mov	r3, r2
+	mov	r2, r9
+
+	@ Load the IV and counter portion.
+	ldr	r7, [r8, #12]
+	vld1.8	{q7}, [r8]
+
+	bl	_vpaes_preheat
+	rev	r7, r7		@ The counter is big-endian.
+
+.Lctr32_loop:
+	vmov	q0, q7
+	vld1.8	{q6}, [r0]!		@ .Load input ahead of time
+	bl	_vpaes_encrypt_core
+	veor	q0, q0, q6		@ XOR input and result
+	vst1.8	{q0}, [r1]!
+	subs	r3, r3, #1
+	@ Update the counter.
+	add	r7, r7, #1
+	rev	r9, r7
+	vmov.32	d15[1], r9
+	bne	.Lctr32_loop
+
+.Lctr32_done:
+	vldmia	sp!, {d8,d9,d10,d11,d12,d13,d14,d15}
+	ldmia	sp!, {r7,r8,r9,r10,r11, pc}	@ return
+.size	vpaes_ctr32_encrypt_blocks,.-vpaes_ctr32_encrypt_blocks
+#endif
+#endif  // !OPENSSL_NO_ASM