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
20 extern int aegis128_have_aes_insn;
22 void *memcpy(void *dest, const void *src, size_t n);
24 struct aegis128_state {
28 extern const uint8_t crypto_aes_sbox[];
30 static struct aegis128_state aegis128_load_state_neon(const void *state)
32 return (struct aegis128_state){ {
41 static void aegis128_save_state_neon(struct aegis128_state st, void *state)
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]);
50 static inline __attribute__((always_inline))
51 uint8x16_t aegis_aes_round(uint8x16_t w)
56 if (!__builtin_expect(aegis128_have_aes_insn, 1)) {
57 static const uint8_t shift_rows[] = {
58 0x0, 0x5, 0xa, 0xf, 0x4, 0x9, 0xe, 0x3,
59 0x8, 0xd, 0x2, 0x7, 0xc, 0x1, 0x6, 0xb,
61 static const uint8_t ror32by8[] = {
62 0x1, 0x2, 0x3, 0x0, 0x5, 0x6, 0x7, 0x4,
63 0x9, 0xa, 0xb, 0x8, 0xd, 0xe, 0xf, 0xc,
68 w = vqtbl1q_u8(w, vld1q_u8(shift_rows));
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);
77 asm("tbl %0.16b, {v16.16b-v19.16b}, %1.16b" : "=w"(v) : "w"(w));
79 asm("tbx %0.16b, {v20.16b-v23.16b}, %1.16b" : "+w"(v) : "w"(w));
81 asm("tbx %0.16b, {v24.16b-v27.16b}, %1.16b" : "+w"(v) : "w"(w));
83 asm("tbx %0.16b, {v28.16b-v31.16b}, %1.16b" : "+w"(v) : "w"(w));
87 w = (v << 1) ^ (uint8x16_t)(((int8x16_t)v >> 7) & 0x1b);
88 w ^= (uint8x16_t)vrev32q_u16((uint16x8_t)v);
89 w ^= vqtbl1q_u8(v ^ w, vld1q_u8(ror32by8));
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.
101 asm(AES_ROUND : "+w"(w) : "w"(z));
105 static inline __attribute__((always_inline))
106 struct aegis128_state aegis128_update_neon(struct aegis128_state st,
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]);
119 static inline __attribute__((always_inline))
120 void preload_sbox(void)
122 if (!IS_ENABLED(CONFIG_ARM64) ||
123 !IS_ENABLED(CONFIG_CC_IS_GCC) ||
124 __builtin_expect(aegis128_have_aes_insn, 1))
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));
134 void crypto_aegis128_init_neon(void *state, const void *key, const void *iv)
136 static const uint8_t const0[] = {
137 0x00, 0x01, 0x01, 0x02, 0x03, 0x05, 0x08, 0x0d,
138 0x15, 0x22, 0x37, 0x59, 0x90, 0xe9, 0x79, 0x62,
140 static const uint8_t const1[] = {
141 0xdb, 0x3d, 0x18, 0x55, 0x6d, 0xc2, 0x2f, 0xf1,
142 0x20, 0x11, 0x31, 0x42, 0x73, 0xb5, 0x28, 0xdd,
144 uint8x16_t k = vld1q_u8(key);
145 uint8x16_t kiv = k ^ vld1q_u8(iv);
146 struct aegis128_state st = {{
150 k ^ vld1q_u8(const0),
151 k ^ vld1q_u8(const1),
157 for (i = 0; i < 5; i++) {
158 st = aegis128_update_neon(st, k);
159 st = aegis128_update_neon(st, kiv);
161 aegis128_save_state_neon(st, state);
164 void crypto_aegis128_update_neon(void *state, const void *msg)
166 struct aegis128_state st = aegis128_load_state_neon(state);
170 st = aegis128_update_neon(st, vld1q_u8(msg));
172 aegis128_save_state_neon(st, state);
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.
181 static uint8x16_t vqtbl1q_u8(uint8x16_t a, uint8x16_t b)
188 return vcombine_u8(vtbl2_u8(__a.pair, vget_low_u8(b)),
189 vtbl2_u8(__a.pair, vget_high_u8(b)));
192 static uint8x16_t vqtbx1q_u8(uint8x16_t v, uint8x16_t a, uint8x16_t b)
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)));
203 static int8_t vminvq_s8(int8x16_t v)
205 int8x8_t s = vpmin_s8(vget_low_s8(v), vget_high_s8(v));
211 return vget_lane_s8(s, 0);
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,
221 void crypto_aegis128_encrypt_chunk_neon(void *state, void *dst, const void *src,
224 struct aegis128_state st = aegis128_load_state_neon(state);
225 const int short_input = size < AEGIS_BLOCK_SIZE;
230 while (size >= AEGIS_BLOCK_SIZE) {
231 uint8x16_t s = st.v[1] ^ (st.v[2] & st.v[3]) ^ st.v[4];
234 st = aegis128_update_neon(st, msg);
238 size -= AEGIS_BLOCK_SIZE;
239 src += AEGIS_BLOCK_SIZE;
240 dst += AEGIS_BLOCK_SIZE;
244 uint8x16_t s = st.v[1] ^ (st.v[2] & st.v[3]) ^ st.v[4];
245 uint8_t buf[AEGIS_BLOCK_SIZE];
246 const void *in = src;
250 if (__builtin_expect(short_input, 0))
251 in = out = memcpy(buf + AEGIS_BLOCK_SIZE - size, src, size);
253 m = vqtbl1q_u8(vld1q_u8(in + size - AEGIS_BLOCK_SIZE),
254 vld1q_u8(permute + 32 - size));
256 st = aegis128_update_neon(st, m);
258 vst1q_u8(out + size - AEGIS_BLOCK_SIZE,
259 vqtbl1q_u8(m ^ s, vld1q_u8(permute + size)));
261 if (__builtin_expect(short_input, 0))
262 memcpy(dst, out, size);
264 vst1q_u8(out - AEGIS_BLOCK_SIZE, msg);
267 aegis128_save_state_neon(st, state);
270 void crypto_aegis128_decrypt_chunk_neon(void *state, void *dst, const void *src,
273 struct aegis128_state st = aegis128_load_state_neon(state);
274 const int short_input = size < AEGIS_BLOCK_SIZE;
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);
284 size -= AEGIS_BLOCK_SIZE;
285 src += AEGIS_BLOCK_SIZE;
286 dst += AEGIS_BLOCK_SIZE;
290 uint8x16_t s = st.v[1] ^ (st.v[2] & st.v[3]) ^ st.v[4];
291 uint8_t buf[AEGIS_BLOCK_SIZE];
292 const void *in = src;
296 if (__builtin_expect(short_input, 0))
297 in = out = memcpy(buf + AEGIS_BLOCK_SIZE - size, src, size);
299 m = s ^ vqtbx1q_u8(s, vld1q_u8(in + size - AEGIS_BLOCK_SIZE),
300 vld1q_u8(permute + 32 - size));
302 st = aegis128_update_neon(st, m);
304 vst1q_u8(out + size - AEGIS_BLOCK_SIZE,
305 vqtbl1q_u8(m, vld1q_u8(permute + size)));
307 if (__builtin_expect(short_input, 0))
308 memcpy(dst, out, size);
310 vst1q_u8(out - AEGIS_BLOCK_SIZE, msg);
313 aegis128_save_state_neon(st, state);
316 int crypto_aegis128_final_neon(void *state, void *tag_xor,
317 unsigned int assoclen,
318 unsigned int cryptlen,
319 unsigned int authsize)
321 struct aegis128_state st = aegis128_load_state_neon(state);
327 v = st.v[3] ^ (uint8x16_t)vcombine_u64(vmov_n_u64(8ULL * assoclen),
328 vmov_n_u64(8ULL * cryptlen));
330 for (i = 0; i < 7; i++)
331 st = aegis128_update_neon(st, v);
333 v = st.v[0] ^ st.v[1] ^ st.v[2] ^ st.v[3] ^ st.v[4];
336 v = vqtbl1q_u8(~vceqq_u8(v, vld1q_u8(tag_xor)),
337 vld1q_u8(permute + authsize));
339 return vminvq_s8((int8x16_t)v);
342 vst1q_u8(tag_xor, v);