1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Copyright (C) 2019 Linaro, Ltd. <ard.biesheuvel@linaro.org>
7 #include <asm/neon-intrinsics.h>
9 #define AES_ROUND "aese %0.16b, %1.16b \n\t aesmc %0.16b, %0.16b"
13 #define AES_ROUND "aese.8 %q0, %q1 \n\t aesmc.8 %q0, %q0"
16 #define AEGIS_BLOCK_SIZE 16
19 #include "aegis-neon.h"
21 extern int aegis128_have_aes_insn;
23 void *memcpy(void *dest, const void *src, size_t n);
25 struct aegis128_state {
29 extern const uint8_t crypto_aes_sbox[];
31 static struct aegis128_state aegis128_load_state_neon(const void *state)
33 return (struct aegis128_state){ {
42 static void aegis128_save_state_neon(struct aegis128_state st, void *state)
44 vst1q_u8(state, st.v[0]);
45 vst1q_u8(state + 16, st.v[1]);
46 vst1q_u8(state + 32, st.v[2]);
47 vst1q_u8(state + 48, st.v[3]);
48 vst1q_u8(state + 64, st.v[4]);
51 static inline __attribute__((always_inline))
52 uint8x16_t aegis_aes_round(uint8x16_t w)
57 if (!__builtin_expect(aegis128_have_aes_insn, 1)) {
58 static const uint8_t shift_rows[] = {
59 0x0, 0x5, 0xa, 0xf, 0x4, 0x9, 0xe, 0x3,
60 0x8, 0xd, 0x2, 0x7, 0xc, 0x1, 0x6, 0xb,
62 static const uint8_t ror32by8[] = {
63 0x1, 0x2, 0x3, 0x0, 0x5, 0x6, 0x7, 0x4,
64 0x9, 0xa, 0xb, 0x8, 0xd, 0xe, 0xf, 0xc,
69 w = vqtbl1q_u8(w, vld1q_u8(shift_rows));
72 #ifndef CONFIG_CC_IS_GCC
73 v = vqtbl4q_u8(vld1q_u8_x4(crypto_aes_sbox), w);
74 v = vqtbx4q_u8(v, vld1q_u8_x4(crypto_aes_sbox + 0x40), w - 0x40);
75 v = vqtbx4q_u8(v, vld1q_u8_x4(crypto_aes_sbox + 0x80), w - 0x80);
76 v = vqtbx4q_u8(v, vld1q_u8_x4(crypto_aes_sbox + 0xc0), w - 0xc0);
78 asm("tbl %0.16b, {v16.16b-v19.16b}, %1.16b" : "=w"(v) : "w"(w));
80 asm("tbx %0.16b, {v20.16b-v23.16b}, %1.16b" : "+w"(v) : "w"(w));
82 asm("tbx %0.16b, {v24.16b-v27.16b}, %1.16b" : "+w"(v) : "w"(w));
84 asm("tbx %0.16b, {v28.16b-v31.16b}, %1.16b" : "+w"(v) : "w"(w));
88 w = (v << 1) ^ (uint8x16_t)(((int8x16_t)v >> 7) & 0x1b);
89 w ^= (uint8x16_t)vrev32q_u16((uint16x8_t)v);
90 w ^= vqtbl1q_u8(v ^ w, vld1q_u8(ror32by8));
97 * We use inline asm here instead of the vaeseq_u8/vaesmcq_u8 intrinsics
98 * to force the compiler to issue the aese/aesmc instructions in pairs.
99 * This is much faster on many cores, where the instruction pair can
100 * execute in a single cycle.
102 asm(AES_ROUND : "+w"(w) : "w"(z));
106 static inline __attribute__((always_inline))
107 struct aegis128_state aegis128_update_neon(struct aegis128_state st,
110 m ^= aegis_aes_round(st.v[4]);
111 st.v[4] ^= aegis_aes_round(st.v[3]);
112 st.v[3] ^= aegis_aes_round(st.v[2]);
113 st.v[2] ^= aegis_aes_round(st.v[1]);
114 st.v[1] ^= aegis_aes_round(st.v[0]);
120 static inline __attribute__((always_inline))
121 void preload_sbox(void)
123 if (!IS_ENABLED(CONFIG_ARM64) ||
124 !IS_ENABLED(CONFIG_CC_IS_GCC) ||
125 __builtin_expect(aegis128_have_aes_insn, 1))
128 asm("ld1 {v16.16b-v19.16b}, [%0], #64 \n\t"
129 "ld1 {v20.16b-v23.16b}, [%0], #64 \n\t"
130 "ld1 {v24.16b-v27.16b}, [%0], #64 \n\t"
131 "ld1 {v28.16b-v31.16b}, [%0] \n\t"
132 :: "r"(crypto_aes_sbox));
135 void crypto_aegis128_init_neon(void *state, const void *key, const void *iv)
137 static const uint8_t const0[] = {
138 0x00, 0x01, 0x01, 0x02, 0x03, 0x05, 0x08, 0x0d,
139 0x15, 0x22, 0x37, 0x59, 0x90, 0xe9, 0x79, 0x62,
141 static const uint8_t const1[] = {
142 0xdb, 0x3d, 0x18, 0x55, 0x6d, 0xc2, 0x2f, 0xf1,
143 0x20, 0x11, 0x31, 0x42, 0x73, 0xb5, 0x28, 0xdd,
145 uint8x16_t k = vld1q_u8(key);
146 uint8x16_t kiv = k ^ vld1q_u8(iv);
147 struct aegis128_state st = {{
151 k ^ vld1q_u8(const0),
152 k ^ vld1q_u8(const1),
158 for (i = 0; i < 5; i++) {
159 st = aegis128_update_neon(st, k);
160 st = aegis128_update_neon(st, kiv);
162 aegis128_save_state_neon(st, state);
165 void crypto_aegis128_update_neon(void *state, const void *msg)
167 struct aegis128_state st = aegis128_load_state_neon(state);
171 st = aegis128_update_neon(st, vld1q_u8(msg));
173 aegis128_save_state_neon(st, state);
178 * AArch32 does not provide these intrinsics natively because it does not
179 * implement the underlying instructions. AArch32 only provides 64-bit
180 * wide vtbl.8/vtbx.8 instruction, so use those instead.
182 static uint8x16_t vqtbl1q_u8(uint8x16_t a, uint8x16_t b)
189 return vcombine_u8(vtbl2_u8(__a.pair, vget_low_u8(b)),
190 vtbl2_u8(__a.pair, vget_high_u8(b)));
193 static uint8x16_t vqtbx1q_u8(uint8x16_t v, uint8x16_t a, uint8x16_t b)
200 return vcombine_u8(vtbx2_u8(vget_low_u8(v), __a.pair, vget_low_u8(b)),
201 vtbx2_u8(vget_high_u8(v), __a.pair, vget_high_u8(b)));
204 static int8_t vminvq_s8(int8x16_t v)
206 int8x8_t s = vpmin_s8(vget_low_s8(v), vget_high_s8(v));
212 return vget_lane_s8(s, 0);
216 static const uint8_t permute[] __aligned(64) = {
217 -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
218 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,
219 -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
222 void crypto_aegis128_encrypt_chunk_neon(void *state, void *dst, const void *src,
225 struct aegis128_state st = aegis128_load_state_neon(state);
226 const int short_input = size < AEGIS_BLOCK_SIZE;
231 while (size >= AEGIS_BLOCK_SIZE) {
232 uint8x16_t s = st.v[1] ^ (st.v[2] & st.v[3]) ^ st.v[4];
235 st = aegis128_update_neon(st, msg);
239 size -= AEGIS_BLOCK_SIZE;
240 src += AEGIS_BLOCK_SIZE;
241 dst += AEGIS_BLOCK_SIZE;
245 uint8x16_t s = st.v[1] ^ (st.v[2] & st.v[3]) ^ st.v[4];
246 uint8_t buf[AEGIS_BLOCK_SIZE];
247 const void *in = src;
251 if (__builtin_expect(short_input, 0))
252 in = out = memcpy(buf + AEGIS_BLOCK_SIZE - size, src, size);
254 m = vqtbl1q_u8(vld1q_u8(in + size - AEGIS_BLOCK_SIZE),
255 vld1q_u8(permute + 32 - size));
257 st = aegis128_update_neon(st, m);
259 vst1q_u8(out + size - AEGIS_BLOCK_SIZE,
260 vqtbl1q_u8(m ^ s, vld1q_u8(permute + size)));
262 if (__builtin_expect(short_input, 0))
263 memcpy(dst, out, size);
265 vst1q_u8(out - AEGIS_BLOCK_SIZE, msg);
268 aegis128_save_state_neon(st, state);
271 void crypto_aegis128_decrypt_chunk_neon(void *state, void *dst, const void *src,
274 struct aegis128_state st = aegis128_load_state_neon(state);
275 const int short_input = size < AEGIS_BLOCK_SIZE;
280 while (size >= AEGIS_BLOCK_SIZE) {
281 msg = vld1q_u8(src) ^ st.v[1] ^ (st.v[2] & st.v[3]) ^ st.v[4];
282 st = aegis128_update_neon(st, msg);
285 size -= AEGIS_BLOCK_SIZE;
286 src += AEGIS_BLOCK_SIZE;
287 dst += AEGIS_BLOCK_SIZE;
291 uint8x16_t s = st.v[1] ^ (st.v[2] & st.v[3]) ^ st.v[4];
292 uint8_t buf[AEGIS_BLOCK_SIZE];
293 const void *in = src;
297 if (__builtin_expect(short_input, 0))
298 in = out = memcpy(buf + AEGIS_BLOCK_SIZE - size, src, size);
300 m = s ^ vqtbx1q_u8(s, vld1q_u8(in + size - AEGIS_BLOCK_SIZE),
301 vld1q_u8(permute + 32 - size));
303 st = aegis128_update_neon(st, m);
305 vst1q_u8(out + size - AEGIS_BLOCK_SIZE,
306 vqtbl1q_u8(m, vld1q_u8(permute + size)));
308 if (__builtin_expect(short_input, 0))
309 memcpy(dst, out, size);
311 vst1q_u8(out - AEGIS_BLOCK_SIZE, msg);
314 aegis128_save_state_neon(st, state);
317 int crypto_aegis128_final_neon(void *state, void *tag_xor,
318 unsigned int assoclen,
319 unsigned int cryptlen,
320 unsigned int authsize)
322 struct aegis128_state st = aegis128_load_state_neon(state);
328 v = st.v[3] ^ (uint8x16_t)vcombine_u64(vmov_n_u64(8ULL * assoclen),
329 vmov_n_u64(8ULL * cryptlen));
331 for (i = 0; i < 7; i++)
332 st = aegis128_update_neon(st, v);
334 v = st.v[0] ^ st.v[1] ^ st.v[2] ^ st.v[3] ^ st.v[4];
337 v = vqtbl1q_u8(~vceqq_u8(v, vld1q_u8(tag_xor)),
338 vld1q_u8(permute + authsize));
340 return vminvq_s8((int8x16_t)v);
343 vst1q_u8(tag_xor, v);