void block(state& s)
{
//ChaCha20 runs 20 rounds, alternating between "column rounds" and
//"diagonal rounds". Each round consists of four quarter-rounds, and
//they are run as follows. Quarter rounds 1-4 are part of a "column"
//round, while 5-8 are part of a "diagonal" round:
const state initial = s;
for (int i = 0; i < 10; ++i) {
quarter_round(s[ 0], s[ 4], s[ 8], s[12]);
quarter_round(s[ 1], s[ 5], s[ 9], s[13]);
quarter_round(s[ 2], s[ 6], s[10], s[14]);
quarter_round(s[ 3], s[ 7], s[11], s[15]);
quarter_round(s[ 0], s[ 5], s[10], s[15]);
quarter_round(s[ 1], s[ 6], s[11], s[12]);
quarter_round(s[ 2], s[ 7], s[ 8], s[13]);
quarter_round(s[ 3], s[ 4], s[ 9], s[14]);
}
// At the end of 20 rounds (or 10 iterations of the above list), we add
// the original input words to the output words, and serialize the
// result by sequencing the words one-by-one in little-endian order.
for (int i = 0; i < 16; ++i) {
s[i] += initial[i];
}
}
extern_c void
crandom_chacha_expand(u_int64_t iv,
u_int64_t ctr,
int nr,
int output_size,
const unsigned char *key_,
unsigned char *output_) {
# if MIGHT_HAVE_SSE2
if (HAVE(SSE2)) {
ssereg *key = (ssereg *)key_;
ssereg *output = (ssereg *)output_;
ssereg a1 = key[0], a2 = a1, aa = a1,
b1 = key[1], b2 = b1, bb = b1,
c1 = {iv, ctr}, c2 = {iv, ctr+1}, cc = c1,
d1 = {0x3320646e61707865ull, 0x6b20657479622d32ull}, d2 = d1, dd = d1,
p = {0, 1};
int i,r;
# if (NEED_XOP)
if (HAVE(XOP)) {
for (i=0; i<output_size; i+=128) {
for (r=nr; r>0; r-=2)
DOUBLE_ROUND(quarter_round_xop);
OUTPUT_FUNCTION;
}
return;
}
# endif
# if (NEED_SSSE3)
if (HAVE(SSSE3)) {
for (i=0; i<output_size; i+=128) {
for (r=nr; r>0; r-=2)
DOUBLE_ROUND(quarter_round_ssse3);
OUTPUT_FUNCTION;
}
return;
}
# endif
# if (NEED_SSE2)
if (HAVE(SSE2)) {
for (i=0; i<output_size; i+=128) {
for (r=nr; r>0; r-=2)
DOUBLE_ROUND(quarter_round_sse2);
OUTPUT_FUNCTION;
}
return;
}
# endif
}
# endif
# if NEED_CONV
{
const u_int32_t *key = (const u_int32_t *)key_;
u_int32_t
x[16],
input[16] = {
key[0], key[1], key[2], key[3],
key[4], key[5], key[6], key[7],
iv, iv>>32, ctr, ctr>>32,
0x61707865, 0x3320646e, 0x79622d32, 0x6b206574
},
*output = (u_int32_t *)output_;
int i, r;
for (i=0; i<output_size; i+= 64) {
for (r=0; r<16; r++) {
x[r] = input[r];
}
for (r=nr; r>0; r-=2) {
quarter_round(&x[0], &x[4], &x[8], &x[12]);
quarter_round(&x[1], &x[5], &x[9], &x[13]);
quarter_round(&x[2], &x[6], &x[10], &x[14]);
quarter_round(&x[3], &x[7], &x[11], &x[15]);
quarter_round(&x[0], &x[5], &x[10], &x[15]);
quarter_round(&x[1], &x[6], &x[11], &x[12]);
quarter_round(&x[2], &x[7], &x[8], &x[13]);
quarter_round(&x[3], &x[4], &x[9], &x[14]);
}
for (r=0; r<16; r++) {
output[r] = x[r] + input[r];
}
output += 16;
input[11] ++;
if (!input[11]) input[12]++;
}
}
