8 INLINE void g(uint32_t *state, size_t a, size_t b, size_t c, size_t d,  in g()  argument
10   state[a] = state[a] + state[b] + x;  in g()
11   state[d] = rotr32(state[d] ^ state[a], 16);  in g()
12   state[c] = state[c] + state[d];  in g()
13   state[b] = rotr32(state[b] ^ state[c], 12);  in g()
14   state[a] = state[a] + state[b] + y;  in g()
15   state[d] = rotr32(state[d] ^ state[a], 8);  in g()
16   state[c] = state[c] + state[d];  in g()
17   state[b] = rotr32(state[b] ^ state[c], 7);  in g()
20 INLINE void round_fn(uint32_t state[16], const uint32_t *msg, size_t round) {  in round_fn()
25   g(state, 0, 4, 8, 12, msg[schedule[0]], msg[schedule[1]]);  in round_fn()
26   g(state, 1, 5, 9, 13, msg[schedule[2]], msg[schedule[3]]);  in round_fn()
27   g(state, 2, 6, 10, 14, msg[schedule[4]], msg[schedule[5]]);  in round_fn()
28   g(state, 3, 7, 11, 15, msg[schedule[6]], msg[schedule[7]]);  in round_fn()
31   g(state, 0, 5, 10, 15, msg[schedule[8]], msg[schedule[9]]);  in round_fn()
32   g(state, 1, 6, 11, 12, msg[schedule[10]], msg[schedule[11]]);  in round_fn()
33   g(state, 2, 7, 8, 13, msg[schedule[12]], msg[schedule[13]]);  in round_fn()
34   g(state, 3, 4, 9, 14, msg[schedule[14]], msg[schedule[15]]);  in round_fn()
37 INLINE void compress_pre(uint32_t state[16], const uint32_t cv[8],  in compress_pre()
58   state[0] = cv[0];  in compress_pre()
59   state[1] = cv[1];  in compress_pre()
60   state[2] = cv[2];  in compress_pre()
61   state[3] = cv[3];  in compress_pre()
62   state[4] = cv[4];  in compress_pre()
63   state[5] = cv[5];  in compress_pre()
64   state[6] = cv[6];  in compress_pre()
65   state[7] = cv[7];  in compress_pre()
66   state[8] = IV[0];  in compress_pre()
67   state[9] = IV[1];  in compress_pre()
68   state[10] = IV[2];  in compress_pre()
69   state[11] = IV[3];  in compress_pre()
70   state[12] = counter_low(counter);  in compress_pre()
71   state[13] = counter_high(counter);  in compress_pre()
72   state[14] = (uint32_t)block_len;  in compress_pre()
73   state[15] = (uint32_t)flags;  in compress_pre()
75   round_fn(state, &block_words[0], 0);  in compress_pre()
76   round_fn(state, &block_words[0], 1);  in compress_pre()
77   round_fn(state, &block_words[0], 2);  in compress_pre()
78   round_fn(state, &block_words[0], 3);  in compress_pre()
79   round_fn(state, &block_words[0], 4);  in compress_pre()
80   round_fn(state, &block_words[0], 5);  in compress_pre()
81   round_fn(state, &block_words[0], 6);  in compress_pre()
88   uint32_t state[16];  in blake3_compress_in_place_portable()  local
89   compress_pre(state, cv, block, block_len, counter, flags);  in blake3_compress_in_place_portable()
90   cv[0] = state[0] ^ state[8];  in blake3_compress_in_place_portable()
91   cv[1] = state[1] ^ state[9];  in blake3_compress_in_place_portable()
92   cv[2] = state[2] ^ state[10];  in blake3_compress_in_place_portable()
93   cv[3] = state[3] ^ state[11];  in blake3_compress_in_place_portable()
94   cv[4] = state[4] ^ state[12];  in blake3_compress_in_place_portable()
95   cv[5] = state[5] ^ state[13];  in blake3_compress_in_place_portable()
96   cv[6] = state[6] ^ state[14];  in blake3_compress_in_place_portable()
97   cv[7] = state[7] ^ state[15];  in blake3_compress_in_place_portable()
104   uint32_t state[16];  in blake3_compress_xof_portable()  local
105   compress_pre(state, cv, block, block_len, counter, flags);  in blake3_compress_xof_portable()
107   store32(&out[0 * 4], state[0] ^ state[8]);  in blake3_compress_xof_portable()
108   store32(&out[1 * 4], state[1] ^ state[9]);  in blake3_compress_xof_portable()
109   store32(&out[2 * 4], state[2] ^ state[10]);  in blake3_compress_xof_portable()
110   store32(&out[3 * 4], state[3] ^ state[11]);  in blake3_compress_xof_portable()
111   store32(&out[4 * 4], state[4] ^ state[12]);  in blake3_compress_xof_portable()
112   store32(&out[5 * 4], state[5] ^ state[13]);  in blake3_compress_xof_portable()
113   store32(&out[6 * 4], state[6] ^ state[14]);  in blake3_compress_xof_portable()
114   store32(&out[7 * 4], state[7] ^ state[15]);  in blake3_compress_xof_portable()
115   store32(&out[8 * 4], state[8] ^ cv[0]);  in blake3_compress_xof_portable()
116   store32(&out[9 * 4], state[9] ^ cv[1]);  in blake3_compress_xof_portable()
117   store32(&out[10 * 4], state[10] ^ cv[2]);  in blake3_compress_xof_portable()
118   store32(&out[11 * 4], state[11] ^ cv[3]);  in blake3_compress_xof_portable()
119   store32(&out[12 * 4], state[12] ^ cv[4]);  in blake3_compress_xof_portable()
120   store32(&out[13 * 4], state[13] ^ cv[5]);  in blake3_compress_xof_portable()
121   store32(&out[14 * 4], state[14] ^ cv[6]);  in blake3_compress_xof_portable()
122   store32(&out[15 * 4], state[15] ^ cv[7]);  in blake3_compress_xof_portable()