33 static inline void g(uint32_t *state, size_t a, size_t b, size_t c, size_t d,  in g()  argument
36 	state[a] = state[a] + state[b] + x;  in g()
37 	state[d] = rotr32(state[d] ^ state[a], 16);  in g()
38 	state[c] = state[c] + state[d];  in g()
39 	state[b] = rotr32(state[b] ^ state[c], 12);  in g()
40 	state[a] = state[a] + state[b] + y;  in g()
41 	state[d] = rotr32(state[d] ^ state[a], 8);  in g()
42 	state[c] = state[c] + state[d];  in g()
43 	state[b] = rotr32(state[b] ^ state[c], 7);  in g()
46 static inline void round_fn(uint32_t state[16], const uint32_t *msg,  in round_fn()
53 	g(state, 0, 4, 8, 12, msg[schedule[0]], msg[schedule[1]]);  in round_fn()
54 	g(state, 1, 5, 9, 13, msg[schedule[2]], msg[schedule[3]]);  in round_fn()
55 	g(state, 2, 6, 10, 14, msg[schedule[4]], msg[schedule[5]]);  in round_fn()
56 	g(state, 3, 7, 11, 15, msg[schedule[6]], msg[schedule[7]]);  in round_fn()
59 	g(state, 0, 5, 10, 15, msg[schedule[8]], msg[schedule[9]]);  in round_fn()
60 	g(state, 1, 6, 11, 12, msg[schedule[10]], msg[schedule[11]]);  in round_fn()
61 	g(state, 2, 7, 8, 13, msg[schedule[12]], msg[schedule[13]]);  in round_fn()
62 	g(state, 3, 4, 9, 14, msg[schedule[14]], msg[schedule[15]]);  in round_fn()
65 static inline void compress_pre(uint32_t state[16], const uint32_t cv[8],  in compress_pre()
87 	state[0] = cv[0];  in compress_pre()
88 	state[1] = cv[1];  in compress_pre()
89 	state[2] = cv[2];  in compress_pre()
90 	state[3] = cv[3];  in compress_pre()
91 	state[4] = cv[4];  in compress_pre()
92 	state[5] = cv[5];  in compress_pre()
93 	state[6] = cv[6];  in compress_pre()
94 	state[7] = cv[7];  in compress_pre()
95 	state[8] = BLAKE3_IV[0];  in compress_pre()
96 	state[9] = BLAKE3_IV[1];  in compress_pre()
97 	state[10] = BLAKE3_IV[2];  in compress_pre()
98 	state[11] = BLAKE3_IV[3];  in compress_pre()
99 	state[12] = counter_low(counter);  in compress_pre()
100 	state[13] = counter_high(counter);  in compress_pre()
101 	state[14] = (uint32_t)block_len;  in compress_pre()
102 	state[15] = (uint32_t)flags;  in compress_pre()
104 	round_fn(state, &block_words[0], 0);  in compress_pre()
105 	round_fn(state, &block_words[0], 1);  in compress_pre()
106 	round_fn(state, &block_words[0], 2);  in compress_pre()
107 	round_fn(state, &block_words[0], 3);  in compress_pre()
108 	round_fn(state, &block_words[0], 4);  in compress_pre()
109 	round_fn(state, &block_words[0], 5);  in compress_pre()
110 	round_fn(state, &block_words[0], 6);  in compress_pre()
117 	uint32_t state[16];  in blake3_compress_in_place_generic()  local
118 	compress_pre(state, cv, block, block_len, counter, flags);  in blake3_compress_in_place_generic()
119 	cv[0] = state[0] ^ state[8];  in blake3_compress_in_place_generic()
120 	cv[1] = state[1] ^ state[9];  in blake3_compress_in_place_generic()
121 	cv[2] = state[2] ^ state[10];  in blake3_compress_in_place_generic()
122 	cv[3] = state[3] ^ state[11];  in blake3_compress_in_place_generic()
123 	cv[4] = state[4] ^ state[12];  in blake3_compress_in_place_generic()
124 	cv[5] = state[5] ^ state[13];  in blake3_compress_in_place_generic()
125 	cv[6] = state[6] ^ state[14];  in blake3_compress_in_place_generic()
126 	cv[7] = state[7] ^ state[15];  in blake3_compress_in_place_generic()
153 	uint32_t state[16];  in blake3_compress_xof_generic()  local
154 	compress_pre(state, cv, block, block_len, counter, flags);  in blake3_compress_xof_generic()
156 	store32(&out[0 * 4], state[0] ^ state[8]);  in blake3_compress_xof_generic()
157 	store32(&out[1 * 4], state[1] ^ state[9]);  in blake3_compress_xof_generic()
158 	store32(&out[2 * 4], state[2] ^ state[10]);  in blake3_compress_xof_generic()
159 	store32(&out[3 * 4], state[3] ^ state[11]);  in blake3_compress_xof_generic()
160 	store32(&out[4 * 4], state[4] ^ state[12]);  in blake3_compress_xof_generic()
161 	store32(&out[5 * 4], state[5] ^ state[13]);  in blake3_compress_xof_generic()
162 	store32(&out[6 * 4], state[6] ^ state[14]);  in blake3_compress_xof_generic()
163 	store32(&out[7 * 4], state[7] ^ state[15]);  in blake3_compress_xof_generic()
164 	store32(&out[8 * 4], state[8] ^ cv[0]);  in blake3_compress_xof_generic()
165 	store32(&out[9 * 4], state[9] ^ cv[1]);  in blake3_compress_xof_generic()
166 	store32(&out[10 * 4], state[10] ^ cv[2]);  in blake3_compress_xof_generic()
167 	store32(&out[11 * 4], state[11] ^ cv[3]);  in blake3_compress_xof_generic()
168 	store32(&out[12 * 4], state[12] ^ cv[4]);  in blake3_compress_xof_generic()
169 	store32(&out[13 * 4], state[13] ^ cv[5]);  in blake3_compress_xof_generic()
170 	store32(&out[14 * 4], state[14] ^ cv[6]);  in blake3_compress_xof_generic()
171 	store32(&out[15 * 4], state[15] ^ cv[7]);  in blake3_compress_xof_generic()