[linux.git] / crypto / aegis128-neon-inner.c

// SPDX-License-Identifier: GPL-2.0-or-later
/*
 * Copyright (C) 2019 Linaro, Ltd. <[email protected]>
 */

#ifdef CONFIG_ARM64
#include <asm/neon-intrinsics.h>

#define AES_ROUND	"aese %0.16b, %1.16b \n\t aesmc %0.16b, %0.16b"
#else
#include <arm_neon.h>

#define AES_ROUND	"aese.8 %q0, %q1 \n\t aesmc.8 %q0, %q0"
#endif

#define AEGIS_BLOCK_SIZE	16

#include <stddef.h>

extern int aegis128_have_aes_insn;

void *memcpy(void *dest, const void *src, size_t n);

struct aegis128_state {
	uint8x16_t v[5];
};

extern const uint8_t crypto_aes_sbox[];

static struct aegis128_state aegis128_load_state_neon(const void *state)
{
	return (struct aegis128_state){ {
		vld1q_u8(state),
		vld1q_u8(state + 16),
		vld1q_u8(state + 32),
		vld1q_u8(state + 48),
		vld1q_u8(state + 64)
	} };
}

static void aegis128_save_state_neon(struct aegis128_state st, void *state)
{
	vst1q_u8(state, st.v[0]);
	vst1q_u8(state + 16, st.v[1]);
	vst1q_u8(state + 32, st.v[2]);
	vst1q_u8(state + 48, st.v[3]);
	vst1q_u8(state + 64, st.v[4]);
}

static inline __attribute__((always_inline))
uint8x16_t aegis_aes_round(uint8x16_t w)
{
	uint8x16_t z = {};

#ifdef CONFIG_ARM64
	if (!__builtin_expect(aegis128_have_aes_insn, 1)) {
		static const uint8_t shift_rows[] = {
			0x0, 0x5, 0xa, 0xf, 0x4, 0x9, 0xe, 0x3,
			0x8, 0xd, 0x2, 0x7, 0xc, 0x1, 0x6, 0xb,
		};
		static const uint8_t ror32by8[] = {
			0x1, 0x2, 0x3, 0x0, 0x5, 0x6, 0x7, 0x4,
			0x9, 0xa, 0xb, 0x8, 0xd, 0xe, 0xf, 0xc,
		};
		uint8x16_t v;

		// shift rows
		w = vqtbl1q_u8(w, vld1q_u8(shift_rows));

		// sub bytes
#ifndef CONFIG_CC_IS_GCC
		v = vqtbl4q_u8(vld1q_u8_x4(crypto_aes_sbox), w);
		v = vqtbx4q_u8(v, vld1q_u8_x4(crypto_aes_sbox + 0x40), w - 0x40);
		v = vqtbx4q_u8(v, vld1q_u8_x4(crypto_aes_sbox + 0x80), w - 0x80);
		v = vqtbx4q_u8(v, vld1q_u8_x4(crypto_aes_sbox + 0xc0), w - 0xc0);
#else
		asm("tbl %0.16b, {v16.16b-v19.16b}, %1.16b" : "=w"(v) : "w"(w));
		w -= 0x40;
		asm("tbx %0.16b, {v20.16b-v23.16b}, %1.16b" : "+w"(v) : "w"(w));
		w -= 0x40;
		asm("tbx %0.16b, {v24.16b-v27.16b}, %1.16b" : "+w"(v) : "w"(w));
		w -= 0x40;
		asm("tbx %0.16b, {v28.16b-v31.16b}, %1.16b" : "+w"(v) : "w"(w));
#endif

		// mix columns
		w = (v << 1) ^ (uint8x16_t)(((int8x16_t)v >> 7) & 0x1b);
		w ^= (uint8x16_t)vrev32q_u16((uint16x8_t)v);
		w ^= vqtbl1q_u8(v ^ w, vld1q_u8(ror32by8));

		return w;
	}
#endif

	/*
	 * We use inline asm here instead of the vaeseq_u8/vaesmcq_u8 intrinsics
	 * to force the compiler to issue the aese/aesmc instructions in pairs.
	 * This is much faster on many cores, where the instruction pair can
	 * execute in a single cycle.
	 */
	asm(AES_ROUND : "+w"(w) : "w"(z));
	return w;
}

static inline __attribute__((always_inline))
struct aegis128_state aegis128_update_neon(struct aegis128_state st,
					   uint8x16_t m)
{
	m       ^= aegis_aes_round(st.v[4]);
	st.v[4] ^= aegis_aes_round(st.v[3]);
	st.v[3] ^= aegis_aes_round(st.v[2]);
	st.v[2] ^= aegis_aes_round(st.v[1]);
	st.v[1] ^= aegis_aes_round(st.v[0]);
	st.v[0] ^= m;

	return st;
}

static inline __attribute__((always_inline))
void preload_sbox(void)
{
	if (!IS_ENABLED(CONFIG_ARM64) ||
	    !IS_ENABLED(CONFIG_CC_IS_GCC) ||
	    __builtin_expect(aegis128_have_aes_insn, 1))
		return;

	asm("ld1	{v16.16b-v19.16b}, [%0], #64	\n\t"
	    "ld1	{v20.16b-v23.16b}, [%0], #64	\n\t"
	    "ld1	{v24.16b-v27.16b}, [%0], #64	\n\t"
	    "ld1	{v28.16b-v31.16b}, [%0]		\n\t"
	    :: "r"(crypto_aes_sbox));
}

void crypto_aegis128_init_neon(void *state, const void *key, const void *iv)
{
	static const uint8_t const0[] = {
		0x00, 0x01, 0x01, 0x02, 0x03, 0x05, 0x08, 0x0d,
		0x15, 0x22, 0x37, 0x59, 0x90, 0xe9, 0x79, 0x62,
	};
	static const uint8_t const1[] = {
		0xdb, 0x3d, 0x18, 0x55, 0x6d, 0xc2, 0x2f, 0xf1,
		0x20, 0x11, 0x31, 0x42, 0x73, 0xb5, 0x28, 0xdd,
	};
	uint8x16_t k = vld1q_u8(key);
	uint8x16_t kiv = k ^ vld1q_u8(iv);
	struct aegis128_state st = {{
		kiv,
		vld1q_u8(const1),
		vld1q_u8(const0),
		k ^ vld1q_u8(const0),
		k ^ vld1q_u8(const1),
	}};
	int i;

	preload_sbox();

	for (i = 0; i < 5; i++) {
		st = aegis128_update_neon(st, k);
		st = aegis128_update_neon(st, kiv);
	}
	aegis128_save_state_neon(st, state);
}

void crypto_aegis128_update_neon(void *state, const void *msg)
{
	struct aegis128_state st = aegis128_load_state_neon(state);

	preload_sbox();

	st = aegis128_update_neon(st, vld1q_u8(msg));

	aegis128_save_state_neon(st, state);
}

#ifdef CONFIG_ARM
/*
 * AArch32 does not provide these intrinsics natively because it does not
 * implement the underlying instructions. AArch32 only provides 64-bit
 * wide vtbl.8/vtbx.8 instruction, so use those instead.
 */
static uint8x16_t vqtbl1q_u8(uint8x16_t a, uint8x16_t b)
{
	union {
		uint8x16_t	val;
		uint8x8x2_t	pair;
	} __a = { a };

	return vcombine_u8(vtbl2_u8(__a.pair, vget_low_u8(b)),
			   vtbl2_u8(__a.pair, vget_high_u8(b)));
}

static uint8x16_t vqtbx1q_u8(uint8x16_t v, uint8x16_t a, uint8x16_t b)
{
	union {
		uint8x16_t	val;
		uint8x8x2_t	pair;
	} __a = { a };

	return vcombine_u8(vtbx2_u8(vget_low_u8(v), __a.pair, vget_low_u8(b)),
			   vtbx2_u8(vget_high_u8(v), __a.pair, vget_high_u8(b)));
}

static int8_t vminvq_s8(int8x16_t v)
{
	int8x8_t s = vpmin_s8(vget_low_s8(v), vget_high_s8(v));

	s = vpmin_s8(s, s);
	s = vpmin_s8(s, s);
	s = vpmin_s8(s, s);

	return vget_lane_s8(s, 0);
}
#endif

static const uint8_t permute[] __aligned(64) = {
	-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
	 0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15,
	-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
};

void crypto_aegis128_encrypt_chunk_neon(void *state, void *dst, const void *src,
					unsigned int size)
{
	struct aegis128_state st = aegis128_load_state_neon(state);
	const int short_input = size < AEGIS_BLOCK_SIZE;
	uint8x16_t msg;

	preload_sbox();

	while (size >= AEGIS_BLOCK_SIZE) {
		uint8x16_t s = st.v[1] ^ (st.v[2] & st.v[3]) ^ st.v[4];

		msg = vld1q_u8(src);
		st = aegis128_update_neon(st, msg);
		msg ^= s;
		vst1q_u8(dst, msg);

		size -= AEGIS_BLOCK_SIZE;
		src += AEGIS_BLOCK_SIZE;
		dst += AEGIS_BLOCK_SIZE;
	}

	if (size > 0) {
		uint8x16_t s = st.v[1] ^ (st.v[2] & st.v[3]) ^ st.v[4];
		uint8_t buf[AEGIS_BLOCK_SIZE];
		const void *in = src;
		void *out = dst;
		uint8x16_t m;

		if (__builtin_expect(short_input, 0))
			in = out = memcpy(buf + AEGIS_BLOCK_SIZE - size, src, size);

		m = vqtbl1q_u8(vld1q_u8(in + size - AEGIS_BLOCK_SIZE),
			       vld1q_u8(permute + 32 - size));

		st = aegis128_update_neon(st, m);

		vst1q_u8(out + size - AEGIS_BLOCK_SIZE,
			 vqtbl1q_u8(m ^ s, vld1q_u8(permute + size)));

		if (__builtin_expect(short_input, 0))
			memcpy(dst, out, size);
		else
			vst1q_u8(out - AEGIS_BLOCK_SIZE, msg);
	}

	aegis128_save_state_neon(st, state);
}

void crypto_aegis128_decrypt_chunk_neon(void *state, void *dst, const void *src,
					unsigned int size)
{
	struct aegis128_state st = aegis128_load_state_neon(state);
	const int short_input = size < AEGIS_BLOCK_SIZE;
	uint8x16_t msg;

	preload_sbox();

	while (size >= AEGIS_BLOCK_SIZE) {
		msg = vld1q_u8(src) ^ st.v[1] ^ (st.v[2] & st.v[3]) ^ st.v[4];
		st = aegis128_update_neon(st, msg);
		vst1q_u8(dst, msg);

		size -= AEGIS_BLOCK_SIZE;
		src += AEGIS_BLOCK_SIZE;
		dst += AEGIS_BLOCK_SIZE;
	}

	if (size > 0) {
		uint8x16_t s = st.v[1] ^ (st.v[2] & st.v[3]) ^ st.v[4];
		uint8_t buf[AEGIS_BLOCK_SIZE];
		const void *in = src;
		void *out = dst;
		uint8x16_t m;

		if (__builtin_expect(short_input, 0))
			in = out = memcpy(buf + AEGIS_BLOCK_SIZE - size, src, size);

		m = s ^ vqtbx1q_u8(s, vld1q_u8(in + size - AEGIS_BLOCK_SIZE),
				   vld1q_u8(permute + 32 - size));

		st = aegis128_update_neon(st, m);

		vst1q_u8(out + size - AEGIS_BLOCK_SIZE,
			 vqtbl1q_u8(m, vld1q_u8(permute + size)));

		if (__builtin_expect(short_input, 0))
			memcpy(dst, out, size);
		else
			vst1q_u8(out - AEGIS_BLOCK_SIZE, msg);
	}

	aegis128_save_state_neon(st, state);
}

int crypto_aegis128_final_neon(void *state, void *tag_xor,
			       unsigned int assoclen,
			       unsigned int cryptlen,
			       unsigned int authsize)
{
	struct aegis128_state st = aegis128_load_state_neon(state);
	uint8x16_t v;
	int i;

	preload_sbox();

	v = st.v[3] ^ (uint8x16_t)vcombine_u64(vmov_n_u64(8ULL * assoclen),
					       vmov_n_u64(8ULL * cryptlen));

	for (i = 0; i < 7; i++)
		st = aegis128_update_neon(st, v);

	v = st.v[0] ^ st.v[1] ^ st.v[2] ^ st.v[3] ^ st.v[4];

	if (authsize > 0) {
		v = vqtbl1q_u8(~vceqq_u8(v, vld1q_u8(tag_xor)),
			       vld1q_u8(permute + authsize));

		return vminvq_s8((int8x16_t)v);
	}

	vst1q_u8(tag_xor, v);
	return 0;
}
Commit	Line	Data
a4397635 AB	1	// SPDX-License-Identifier: GPL-2.0-or-later
	2	/*
	3	* Copyright (C) 2019 Linaro, Ltd. <[email protected]>
	4	*/
	5
	6	#ifdef CONFIG_ARM64
	7	#include <asm/neon-intrinsics.h>
	8
	9	#define AES_ROUND "aese %0.16b, %1.16b \n\t aesmc %0.16b, %0.16b"
	10	#else
	11	#include <arm_neon.h>
	12
	13	#define AES_ROUND "aese.8 %q0, %q1 \n\t aesmc.8 %q0, %q0"
	14	#endif
	15
	16	#define AEGIS_BLOCK_SIZE 16
	17
	18	#include <stddef.h>
	19
19842963 AB	20	extern int aegis128_have_aes_insn;
19842963 AB	21
a4397635	22	void memcpy(void dest, const void *src, size_t n);
a4397635 AB	23
	24	struct aegis128_state {
	25	uint8x16_t v[5];
	26	};
	27
389139b3	28	extern const uint8_t crypto_aes_sbox[];
19842963	29
a4397635 AB	30	static struct aegis128_state aegis128_load_state_neon(const void *state)
	31	{
	32	return (struct aegis128_state){ {
	33	vld1q_u8(state),
	34	vld1q_u8(state + 16),
	35	vld1q_u8(state + 32),
	36	vld1q_u8(state + 48),
	37	vld1q_u8(state + 64)
	38	} };
	39	}
	40
	41	static void aegis128_save_state_neon(struct aegis128_state st, void *state)
	42	{
	43	vst1q_u8(state, st.v[0]);
	44	vst1q_u8(state + 16, st.v[1]);
	45	vst1q_u8(state + 32, st.v[2]);
	46	vst1q_u8(state + 48, st.v[3]);
	47	vst1q_u8(state + 64, st.v[4]);
	48	}
	49
	50	static inline __attribute__((always_inline))
	51	uint8x16_t aegis_aes_round(uint8x16_t w)
	52	{
	53	uint8x16_t z = {};
	54
19842963 AB	55	#ifdef CONFIG_ARM64
19842963 AB	56	if (!__builtin_expect(aegis128_have_aes_insn, 1)) {
389139b3	57	static const uint8_t shift_rows[] = {
19842963 AB	58	0x0, 0x5, 0xa, 0xf, 0x4, 0x9, 0xe, 0x3,
	59	0x8, 0xd, 0x2, 0x7, 0xc, 0x1, 0x6, 0xb,
	60	};
389139b3	61	static const uint8_t ror32by8[] = {
19842963 AB	62	0x1, 0x2, 0x3, 0x0, 0x5, 0x6, 0x7, 0x4,
	63	0x9, 0xa, 0xb, 0x8, 0xd, 0xe, 0xf, 0xc,
	64	};
	65	uint8x16_t v;
	66
	67	// shift rows
389139b3	68	w = vqtbl1q_u8(w, vld1q_u8(shift_rows));
19842963 AB	69
19842963 AB	70	// sub bytes
389139b3 AB	71	#ifndef CONFIG_CC_IS_GCC
	72	v = vqtbl4q_u8(vld1q_u8_x4(crypto_aes_sbox), w);
	73	v = vqtbx4q_u8(v, vld1q_u8_x4(crypto_aes_sbox + 0x40), w - 0x40);
	74	v = vqtbx4q_u8(v, vld1q_u8_x4(crypto_aes_sbox + 0x80), w - 0x80);
	75	v = vqtbx4q_u8(v, vld1q_u8_x4(crypto_aes_sbox + 0xc0), w - 0xc0);
	76	#else
	77	asm("tbl %0.16b, {v16.16b-v19.16b}, %1.16b" : "=w"(v) : "w"(w));
	78	w -= 0x40;
	79	asm("tbx %0.16b, {v20.16b-v23.16b}, %1.16b" : "+w"(v) : "w"(w));
	80	w -= 0x40;
	81	asm("tbx %0.16b, {v24.16b-v27.16b}, %1.16b" : "+w"(v) : "w"(w));
	82	w -= 0x40;
	83	asm("tbx %0.16b, {v28.16b-v31.16b}, %1.16b" : "+w"(v) : "w"(w));
	84	#endif
19842963 AB	85
	86	// mix columns
	87	w = (v << 1) ^ (uint8x16_t)(((int8x16_t)v >> 7) & 0x1b);
	88	w ^= (uint8x16_t)vrev32q_u16((uint16x8_t)v);
389139b3	89	w ^= vqtbl1q_u8(v ^ w, vld1q_u8(ror32by8));
19842963 AB	90
	91	return w;
	92	}
	93	#endif
	94
a4397635 AB	95	/*
	96	* We use inline asm here instead of the vaeseq_u8/vaesmcq_u8 intrinsics
	97	* to force the compiler to issue the aese/aesmc instructions in pairs.
	98	* This is much faster on many cores, where the instruction pair can
	99	* execute in a single cycle.
	100	*/
	101	asm(AES_ROUND : "+w"(w) : "w"(z));
	102	return w;
	103	}
	104
	105	static inline __attribute__((always_inline))
	106	struct aegis128_state aegis128_update_neon(struct aegis128_state st,
	107	uint8x16_t m)
	108	{
	109	m ^= aegis_aes_round(st.v[4]);
	110	st.v[4] ^= aegis_aes_round(st.v[3]);
	111	st.v[3] ^= aegis_aes_round(st.v[2]);
	112	st.v[2] ^= aegis_aes_round(st.v[1]);
	113	st.v[1] ^= aegis_aes_round(st.v[0]);
	114	st.v[0] ^= m;
	115
	116	return st;
	117	}
	118
19842963 AB	119	static inline __attribute__((always_inline))
	120	void preload_sbox(void)
	121	{
	122	if (!IS_ENABLED(CONFIG_ARM64) \|\|
	123	!IS_ENABLED(CONFIG_CC_IS_GCC) \|\|
	124	__builtin_expect(aegis128_have_aes_insn, 1))
	125	return;
	126
	127	asm("ld1 {v16.16b-v19.16b}, [%0], #64 \n\t"
	128	"ld1 {v20.16b-v23.16b}, [%0], #64 \n\t"
	129	"ld1 {v24.16b-v27.16b}, [%0], #64 \n\t"
	130	"ld1 {v28.16b-v31.16b}, [%0] \n\t"
	131	:: "r"(crypto_aes_sbox));
	132	}
	133
52828263 AB	134	void crypto_aegis128_init_neon(void state, const void key, const void *iv)
	135	{
	136	static const uint8_t const0[] = {
	137	0x00, 0x01, 0x01, 0x02, 0x03, 0x05, 0x08, 0x0d,
	138	0x15, 0x22, 0x37, 0x59, 0x90, 0xe9, 0x79, 0x62,
	139	};
	140	static const uint8_t const1[] = {
	141	0xdb, 0x3d, 0x18, 0x55, 0x6d, 0xc2, 0x2f, 0xf1,
	142	0x20, 0x11, 0x31, 0x42, 0x73, 0xb5, 0x28, 0xdd,
	143	};
	144	uint8x16_t k = vld1q_u8(key);
	145	uint8x16_t kiv = k ^ vld1q_u8(iv);
	146	struct aegis128_state st = {{
	147	kiv,
	148	vld1q_u8(const1),
	149	vld1q_u8(const0),
	150	k ^ vld1q_u8(const0),
	151	k ^ vld1q_u8(const1),
	152	}};
	153	int i;
	154
	155	preload_sbox();
	156
	157	for (i = 0; i < 5; i++) {
	158	st = aegis128_update_neon(st, k);
	159	st = aegis128_update_neon(st, kiv);
	160	}
	161	aegis128_save_state_neon(st, state);
	162	}
	163
a4397635 AB	164	void crypto_aegis128_update_neon(void state, const void msg)
	165	{
	166	struct aegis128_state st = aegis128_load_state_neon(state);
	167
19842963 AB	168	preload_sbox();
19842963 AB	169
a4397635 AB	170	st = aegis128_update_neon(st, vld1q_u8(msg));
	171
	172	aegis128_save_state_neon(st, state);
	173	}
	174
ad00d41b AB	175	#ifdef CONFIG_ARM
	176	/*
	177	* AArch32 does not provide these intrinsics natively because it does not
	178	* implement the underlying instructions. AArch32 only provides 64-bit
	179	* wide vtbl.8/vtbx.8 instruction, so use those instead.
	180	*/
	181	static uint8x16_t vqtbl1q_u8(uint8x16_t a, uint8x16_t b)
	182	{
	183	union {
	184	uint8x16_t val;
	185	uint8x8x2_t pair;
	186	} __a = { a };
	187
	188	return vcombine_u8(vtbl2_u8(__a.pair, vget_low_u8(b)),
	189	vtbl2_u8(__a.pair, vget_high_u8(b)));
	190	}
	191
	192	static uint8x16_t vqtbx1q_u8(uint8x16_t v, uint8x16_t a, uint8x16_t b)
	193	{
	194	union {
	195	uint8x16_t val;
	196	uint8x8x2_t pair;
	197	} __a = { a };
	198
	199	return vcombine_u8(vtbx2_u8(vget_low_u8(v), __a.pair, vget_low_u8(b)),
	200	vtbx2_u8(vget_high_u8(v), __a.pair, vget_high_u8(b)));
	201	}
97b70180 AB	202
	203	static int8_t vminvq_s8(int8x16_t v)
	204	{
	205	int8x8_t s = vpmin_s8(vget_low_s8(v), vget_high_s8(v));
	206
	207	s = vpmin_s8(s, s);
	208	s = vpmin_s8(s, s);
	209	s = vpmin_s8(s, s);
	210
	211	return vget_lane_s8(s, 0);
	212	}
ad00d41b AB	213	#endif
	214
	215	static const uint8_t permute[] __aligned(64) = {
	216	-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
	217	0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,
	218	-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
	219	};
	220
a4397635 AB	221	void crypto_aegis128_encrypt_chunk_neon(void state, void dst, const void *src,
	222	unsigned int size)
	223	{
	224	struct aegis128_state st = aegis128_load_state_neon(state);
ad00d41b	225	const int short_input = size < AEGIS_BLOCK_SIZE;
a4397635 AB	226	uint8x16_t msg;
a4397635 AB	227
19842963 AB	228	preload_sbox();
19842963 AB	229
a4397635 AB	230	while (size >= AEGIS_BLOCK_SIZE) {
	231	uint8x16_t s = st.v[1] ^ (st.v[2] & st.v[3]) ^ st.v[4];
	232
	233	msg = vld1q_u8(src);
	234	st = aegis128_update_neon(st, msg);
ad00d41b AB	235	msg ^= s;
ad00d41b AB	236	vst1q_u8(dst, msg);
a4397635 AB	237
	238	size -= AEGIS_BLOCK_SIZE;
	239	src += AEGIS_BLOCK_SIZE;
	240	dst += AEGIS_BLOCK_SIZE;
	241	}
	242
	243	if (size > 0) {
	244	uint8x16_t s = st.v[1] ^ (st.v[2] & st.v[3]) ^ st.v[4];
ad00d41b AB	245	uint8_t buf[AEGIS_BLOCK_SIZE];
	246	const void *in = src;
	247	void *out = dst;
	248	uint8x16_t m;
a4397635	249
ad00d41b AB	250	if (__builtin_expect(short_input, 0))
	251	in = out = memcpy(buf + AEGIS_BLOCK_SIZE - size, src, size);
	252
	253	m = vqtbl1q_u8(vld1q_u8(in + size - AEGIS_BLOCK_SIZE),
	254	vld1q_u8(permute + 32 - size));
	255
	256	st = aegis128_update_neon(st, m);
	257
	258	vst1q_u8(out + size - AEGIS_BLOCK_SIZE,
	259	vqtbl1q_u8(m ^ s, vld1q_u8(permute + size)));
	260
	261	if (__builtin_expect(short_input, 0))
	262	memcpy(dst, out, size);
	263	else
	264	vst1q_u8(out - AEGIS_BLOCK_SIZE, msg);
a4397635 AB	265	}
	266
	267	aegis128_save_state_neon(st, state);
	268	}
	269
	270	void crypto_aegis128_decrypt_chunk_neon(void state, void dst, const void *src,
	271	unsigned int size)
	272	{
	273	struct aegis128_state st = aegis128_load_state_neon(state);
ad00d41b	274	const int short_input = size < AEGIS_BLOCK_SIZE;
a4397635 AB	275	uint8x16_t msg;
a4397635 AB	276
19842963 AB	277	preload_sbox();
19842963 AB	278
a4397635 AB	279	while (size >= AEGIS_BLOCK_SIZE) {
	280	msg = vld1q_u8(src) ^ st.v[1] ^ (st.v[2] & st.v[3]) ^ st.v[4];
	281	st = aegis128_update_neon(st, msg);
	282	vst1q_u8(dst, msg);
	283
	284	size -= AEGIS_BLOCK_SIZE;
	285	src += AEGIS_BLOCK_SIZE;
	286	dst += AEGIS_BLOCK_SIZE;
	287	}
	288
	289	if (size > 0) {
	290	uint8x16_t s = st.v[1] ^ (st.v[2] & st.v[3]) ^ st.v[4];
	291	uint8_t buf[AEGIS_BLOCK_SIZE];
ad00d41b AB	292	const void *in = src;
	293	void *out = dst;
	294	uint8x16_t m;
a4397635	295
ad00d41b AB	296	if (__builtin_expect(short_input, 0))
ad00d41b AB	297	in = out = memcpy(buf + AEGIS_BLOCK_SIZE - size, src, size);
a4397635	298
ad00d41b AB	299	m = s ^ vqtbx1q_u8(s, vld1q_u8(in + size - AEGIS_BLOCK_SIZE),
	300	vld1q_u8(permute + 32 - size));
	301
	302	st = aegis128_update_neon(st, m);
	303
	304	vst1q_u8(out + size - AEGIS_BLOCK_SIZE,
	305	vqtbl1q_u8(m, vld1q_u8(permute + size)));
	306
	307	if (__builtin_expect(short_input, 0))
	308	memcpy(dst, out, size);
	309	else
	310	vst1q_u8(out - AEGIS_BLOCK_SIZE, msg);
a4397635 AB	311	}
	312
	313	aegis128_save_state_neon(st, state);
	314	}
52828263	315
97b70180 AB	316	int crypto_aegis128_final_neon(void state, void tag_xor,
	317	unsigned int assoclen,
	318	unsigned int cryptlen,
	319	unsigned int authsize)
52828263 AB	320	{
	321	struct aegis128_state st = aegis128_load_state_neon(state);
	322	uint8x16_t v;
	323	int i;
	324
	325	preload_sbox();
	326
97b70180 AB	327	v = st.v[3] ^ (uint8x16_t)vcombine_u64(vmov_n_u64(8ULL * assoclen),
97b70180 AB	328	vmov_n_u64(8ULL * cryptlen));
52828263 AB	329
	330	for (i = 0; i < 7; i++)
	331	st = aegis128_update_neon(st, v);
	332
97b70180 AB	333	v = st.v[0] ^ st.v[1] ^ st.v[2] ^ st.v[3] ^ st.v[4];
	334
	335	if (authsize > 0) {
	336	v = vqtbl1q_u8(~vceqq_u8(v, vld1q_u8(tag_xor)),
	337	vld1q_u8(permute + authsize));
	338
	339	return vminvq_s8((int8x16_t)v);
	340	}
	341
52828263	342	vst1q_u8(tag_xor, v);
97b70180	343	return 0;
52828263	344	}