crypto: aegis128/neon - optimize tail block handling

index 2a660ac1bc3a..81cc7383e54b 100644
@@ -173,10 +173,46 @@ void crypto_aegis128_update_neon(void *state, const void *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)));
+}
+#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();

@@ -186,7 +222,8 @@ void crypto_aegis128_encrypt_chunk_neon(void *state, void *dst, const void *src,
 
 		msg = vld1q_u8(src);
 		st = aegis128_update_neon(st, msg);
-		vst1q_u8(dst, msg ^ s);
+		msg ^= s;
+		vst1q_u8(dst, msg);
 
 		size -= AEGIS_BLOCK_SIZE;
 		src += AEGIS_BLOCK_SIZE;

@@ -195,13 +232,26 @@ void crypto_aegis128_encrypt_chunk_neon(void *state, void *dst, const void *src,
 
 	if (size > 0) {
 		uint8x16_t s = st.v[1] ^ (st.v[2] & st.v[3]) ^ st.v[4];
-		uint8_t buf[AEGIS_BLOCK_SIZE] = {};
+		uint8_t buf[AEGIS_BLOCK_SIZE];
+		const void *in = src;
+		void *out = dst;
+		uint8x16_t m;
 
-		memcpy(buf, src, size);
-		msg = vld1q_u8(buf);
-		st = aegis128_update_neon(st, msg);
-		vst1q_u8(buf, msg ^ s);
-		memcpy(dst, buf, size);
+		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);

@@ -211,6 +261,7 @@ 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();

@@ -228,14 +279,25 @@ void crypto_aegis128_decrypt_chunk_neon(void *state, void *dst, const void *src,
 	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;
 
-		vst1q_u8(buf, s);
-		memcpy(buf, src, size);
-		msg = vld1q_u8(buf) ^ s;
-		vst1q_u8(buf, msg);
-		memcpy(dst, buf, size);
+		if (__builtin_expect(short_input, 0))
+			in = out = memcpy(buf + AEGIS_BLOCK_SIZE - size, src, size);
 
-		st = aegis128_update_neon(st, msg);
+		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);