#include <stdio.h>
#include <stdint.h>
#include <string.h>
#include <stdlib.h>

uint64_t mixer = 6148914691236517205;//010101...
uint64_t carry = 1234567890123456789;

void stir(uint64_t * state, uint64_t statelen)
{
	
	uint64_t j;
	
	for(j = 0; j < statelen; j++)
	{
		if(state[(j+2)%statelen]>state[(j+3)%statelen])
			carry ^= state[(j+1)%statelen];
		else
			carry ^= ~state[(j+1)%statelen];

		state[j] ^= carry;
		carry += mixer;
	}
}

int PHS(void *out, size_t outlen, 
		const void *in, size_t inlen, 
		const void *salt, size_t saltlen, 
		uint64_t t_cost, uint64_t m_cost)
{
	uint64_t state[256] = {0};
	uint64_t * memstate;
	uint64_t rounds = 4, i;
	uint64_t statelen = 256, j;
	
	memmove(&state[0], in, inlen);
	memmove(&state[(inlen / 8) + 1], salt, saltlen);
	state[statelen - 3] = outlen;
	state[statelen - 2] = inlen;
	state[statelen - 1] = saltlen;
	
	stir(state, statelen * rounds);
		
	if(t_cost > 0)
		stir(state, statelen * t_cost);
		
	if(m_cost > 0)
	{
		memstate = (uint64_t *) malloc(m_cost * statelen);
		
		for(i = 0; i < m_cost; i++)
		{
			for(j = 0; j < statelen; j++)
				memstate[j + (i * statelen)] = state[j];
				
			stir(state, statelen);
		}
			
		for(i = 0; i < m_cost * statelen; i++)
		{
			mixer += memstate[carry % m_cost * statelen];
			stir(state, statelen);
		}
		
		free(memstate);
	}
	//outlen is 32
	for(i = 0; i < outlen; i++)
		memcpy(out + i, &state[i * 8], sizeof(char));
	
	return 0;
}