void blake_small_lastBlock(blake_small_ctx_t *ctx, const void *msg,
uint16_t length_b)
{
while (length_b >= BLAKE_SMALL_BLOCKSIZE) {
blake_small_nextBlock(ctx, msg);
msg = (uint8_t*) msg + BLAKE_SMALL_BLOCKSIZE_B;
length_b -= BLAKE_SMALL_BLOCKSIZE;
}
union {
uint8_t v8[64];
uint32_t v32[16];
} buffer;
uint32_t v[16];
union {
uint64_t v64;
uint32_t v32[2];
} ctr;
ctr.v64 = ctx->counter * 512 + length_b;
memset(buffer.v8, 0, 64);
memcpy(buffer.v8, msg, (length_b + 7) / 8);
buffer.v8[length_b / 8] |= 0x80 >> (length_b & 0x7);
blake_small_changeendian(buffer.v8, buffer.v8);
blake_small_expand(v, ctx);
if (length_b > 512 - 64 - 2) {
v[12] ^= ctr.v32[0];
v[13] ^= ctr.v32[0];
v[14] ^= ctr.v32[1];
v[15] ^= ctr.v32[1];
blake_small_compress(v, buffer.v8);
blake_small_collapse(ctx, v);
memset(buffer.v8, 0, 64 - 8);
blake_small_expand(v, ctx);
} else {
if (length_b) {
v[12] ^= ctr.v32[0];
v[13] ^= ctr.v32[0];
v[14] ^= ctr.v32[1];
v[15] ^= ctr.v32[1];
}
}
if (ctx->appendone)
buffer.v8[64 - 8 - 4] |= 0x01;
buffer.v32[14] = ctr.v32[1];
buffer.v32[15] = ctr.v32[0];
blake_small_compress(v, buffer.v8);
blake_small_collapse(ctx, v);
}
void blake_small_nextBlock(blake_small_ctx_t* ctx, const void* msg){
uint32_t v[16];
uint32_t m[16];
union {
uint64_t v64;
uint32_t v32[2];
}ctr;
blake_small_expand(v,ctx);
ctx->counter++;
ctr.v64 = ctx->counter*512;
v[12] ^= ctr.v32[0];
v[13] ^= ctr.v32[0];
v[14] ^= ctr.v32[1];
v[15] ^= ctr.v32[1];
blake_small_changeendian(m, msg);
blake_small_compress(v, m);
blake_small_collapse(ctx, v);
}
