#include <avr/io.h>
#include <avr/delay.h>

static char buffer[300];

#define OUT0 "sbi %0, 5 \n"  /* pull line to GND */
#define OUT1 "cbi %0, 5 \n"  /* tristate the line */

#define WAIT10 "ldi r16, 1 \n rcall delay_loop \n"
#define WAIT34 "ldi r16, 9 \n rcall delay_loop \n"
#define WAIT10MINUS10 ""
#define WAIT34MINUS10 "ldi r16,5 \n rcall delay_loop \n nop \n nop \n"

int recv_gc()
{
	char* buf_ptr = buffer;

	asm volatile(
		"clr r16 \n"
		"recv_wait_low_initial_loop: \n"
		"inc r16 \n"
		"breq timeout \n"
		"sbic %1, 5 \n"
		"rjmp recv_wait_low_initial_loop \n" // from low to the start of counting: 6 cycles = 0.5us
		"nop \n"
		"nop \n"
		"nop \n"
		"nop \n"


		"recv_loop: \n"
		"recv_low_begin: \n"
		"ldi r16, 128 \n" // 1
		"recv_low_loop: \n"
		"inc r16 \n"      // 1
		"breq timeout \n" // 1 if no timeout
		"sbis %1, 5 \n"   // 1               // von high auf dec: 6
		"rjmp recv_low_loop \n\n" // 2 if executed, 1 if skipped.
		"nop \n"          // to account for the rjmp recv_loop below.
		"nop \n"
		"nop \n"
		"nop \n"

		"recv_high_begin: \n"
		"nop \n" // to account for the ldi in recv_low_begin.
		"recv_high_loop: \n"
		"dec r16 \n" // 1
		"breq timeout \n" // 1 if no timeout
		"sbic %1, 5 \n" // 1                 // von low auf inc: 6
		"rjmp recv_high_loop \n\n" // 2 if executed, 1 if skipped

		"st z+, r16 \n" // 2
		"rjmp recv_loop \n\n" // 2

		"timeout: \n"

		:       "+z" ((unsigned char volatile*) buf_ptr)
		:	"I" (_SFR_IO_ADDR(PINC))
		: "r16", "memory"
		);

		return buf_ptr-buffer;

		// a value of >=128 means "0", <127 means "1" bit.
}

void send_gc(char* bytes, int len)
{
	int k=0;
	for (int i=0; i<len; i++)
		for (int j=0x80; j!=0; j=j>>1, k++)
			buffer[k] = ((bytes[i] & j)!=0)?1:0 ;
	
	len=len*8+1;
	
	asm volatile(
		"push r31 \n"
		"push r30 \n\n"

		"send_loop:   \n"
		"sbiw %2, 1 \n" // 2
		"breq send_done \n" // 1 if not done, 2 if done
		"ld r16, z+ \n" // 2
		"tst r16 \n" // 1
		"brne send_one \n\n" // 1 if zero, 2 if nonzero

		"; otherwise, send zero \n"
		"nop \n" // 1
		OUT0
		WAIT34
		OUT1
		WAIT10MINUS10
		"rjmp send_loop \n\n" // 2

		"send_one: \n"
		OUT0
		WAIT10
		OUT1
		WAIT34MINUS10
		"rjmp send_loop \n\n"

		"delay_loop:\n"
		"; this costs 7 + 3*(r16) cycles \n"
		"dec r16\n"
		"brne delay_loop\n"
		"ret\n\n"

		"send_done: \n"
		"; now send the stop bit and release the line \n"
		"nop \n nop \n nop \n"
		OUT0
		"; instead of WAIT10, do some sensible cleanup work \n"
		"clr r16  ; needed below \n"
		"pop r30 \n"
		"pop r31 \n"
		"nop \n nop \n nop \n nop \n nop \n"
		OUT1
		"; done :) \n\n\n"


		"; now the final thing is to wait for DATA become high again (should be immediately anyway) \n"
		"send_final_loop: \n"
		"inc r16 \n"
		"breq send_timeout \n"
		"sbis %0, 5 \n"
		"rjmp send_final_loop \n"
		"send_timeout: \n"
		
		: 
		:	"I" (_SFR_IO_ADDR(DDRC)), 
			"I" (_SFR_IO_ADDR(PINC)),
			"w" (len),
			"z" ((unsigned char volatile*) buffer)
		: "r16"
		);

}


int main (void)
{
	DDRC=0x00;
	PORTC=0x00;

	DDRB=0xFF;
	DDRD=0x00;
	PORTB=0x55;


	int temp=0;
	int n_received;

	while(1)
	{
		temp++;
		if (!(PIND & 0x08))
		{
			//PORTB=~(1 << ((temp>>9)%6));
			PORTB=~temp;
		}
		else if (!(PIND & 0x10))
		{
			//PORTB=~buffer[4];
			char tmp=0;
			for (int i=0;i<8;i++)
				tmp|= ( (buffer[i]&0x80)?(1<<i):0 );
			PORTB=~tmp;
		}
		else if (!(PIND & 0x20))
		{
			PORTB=~n_received>>3;
		}
		else if (!(PIND & 0x40))
		{
			PORTB=~0x00;
		}

               _delay_ms(1);

	char foo[] = { 0x40, 0x03, 0x02 };

	send_gc(foo, 3);
	n_received=recv_gc();


	}


	return 0; // never reached
}