static inline u_int64_t
getblock64(const u_int64_t *p, int i)
{
const u_int8_t *bytes = (const u_int8_t *)(void *)(uintptr_t)(p + i);
u_int64_t value;
if (ALIGNED64(p)) {
value = p[i];
} else {
#if BYTE_ORDER == BIG_ENDIAN
value =
(((u_int64_t)bytes[0]) << 56) |
(((u_int64_t)bytes[1]) << 48) |
(((u_int64_t)bytes[2]) << 40) |
(((u_int64_t)bytes[3]) << 32) |
(((u_int64_t)bytes[4]) << 24) |
(((u_int64_t)bytes[5]) << 16) |
(((u_int64_t)bytes[6]) << 8) |
((u_int64_t)bytes[7]);
#else /* LITTLE_ENDIAN */
value =
(((u_int64_t)bytes[7]) << 56) |
(((u_int64_t)bytes[6]) << 48) |
(((u_int64_t)bytes[5]) << 40) |
(((u_int64_t)bytes[4]) << 32) |
(((u_int64_t)bytes[3]) << 24) |
(((u_int64_t)bytes[2]) << 16) |
(((u_int64_t)bytes[1]) << 8) |
((u_int64_t)bytes[0]);
#endif /* LITTLE_ENDIAN */
}
return (value);
}
void cryptonite_chacha_generate(uint8_t *dst, cryptonite_chacha_context *ctx, uint32_t bytes)
{
cryptonite_chacha_state *st;
block out;
int i;
if (!bytes)
return;
/* xor the previous buffer first (if any) */
if (ctx->prev_len > 0) {
int to_copy = (ctx->prev_len < bytes) ? ctx->prev_len : bytes;
for (i = 0; i < to_copy; i++)
dst[i] = ctx->prev[ctx->prev_ofs+i];
memset(ctx->prev + ctx->prev_ofs, 0, to_copy);
ctx->prev_len -= to_copy;
ctx->prev_ofs += to_copy;
dst += to_copy;
bytes -= to_copy;
}
if (bytes == 0)
return;
st = &ctx->st;
if (ALIGNED64(dst)) {
/* xor new 64-bytes chunks and store the left over if any */
for (; bytes >= 64; bytes -= 64, dst += 64) {
/* generate new chunk and update state */
chacha_core(ctx->nb_rounds, (block *) dst, st);
st->d[12] += 1;
if (st->d[12] == 0)
st->d[13] += 1;
}
} else {
/* xor new 64-bytes chunks and store the left over if any */
for (; bytes >= 64; bytes -= 64, dst += 64) {
/* generate new chunk and update state */
chacha_core(ctx->nb_rounds, &out, st);
st->d[12] += 1;
if (st->d[12] == 0)
st->d[13] += 1;
for (i = 0; i < 64; ++i)
dst[i] = out.b[i];
}
}
if (bytes > 0) {
/* generate new chunk and update state */
chacha_core(ctx->nb_rounds, &out, st);
st->d[12] += 1;
if (st->d[12] == 0)
st->d[13] += 1;
/* xor as much as needed */
for (i = 0; i < bytes; i++)
dst[i] = out.b[i];
/* copy the left over in the buffer */
ctx->prev_len = 64 - bytes;
ctx->prev_ofs = i;
for (; i < 64; i++)
ctx->prev[i] = out.b[i];
}
}
