#include "opcode.h"
#include <assert.h>
#include <ctype.h>
#include <string.h>
#include <pthread.h>
#define bench 0
#define debug 0
#define IO 1
#define keypoll 0
#if bench
#include <sys/time.h>
#else
#include <curses.h>
#endif

#define THREADS	1
#define BENCH_INST 100000000 << THREADS-1
#define CTRL_ADDR 0xC000
#define TX_ADDR 0xC001
#define RX_ADDR 0xC002
#define STEP_ADDR 0xC010
#define CURSES_BACKSPACE 0x7F

#define setflag(flag, bit) (flag) ? (cpu->ps |= (bit << (thread << 3))) : (cpu->ps &= ~(bit << (thread << 3)))
#define clearflags(thread) {\
			cpu->c[thread] = 0;\
			cpu->z[thread] = 0;\
			cpu->i[thread] = 0;\
			cpu->s[thread] = 0;\
			cpu->v[thread] = 0;\
			cpu->n[thread] = 0;\
}

uint64_t clk[THREADS];	/* Per Thread Clock cycles. */
uint64_t tclk;		/* Total Clock cycles. */
uint64_t inst[THREADS];
uint64_t inss;
uint8_t threads_done = 0;
uint8_t kbd_rdy = 0;
uint8_t kbd_ln = 0;
uint8_t wai = 0;
uint8_t step = 0;
uint8_t irq = 0;
#if !bench
WINDOW *scr;
#endif
pthread_mutex_t mutex = PTHREAD_MUTEX_INITIALIZER;
pthread_mutex_t main_mutex = PTHREAD_MUTEX_INITIALIZER;
pthread_cond_t cond = PTHREAD_COND_INITIALIZER;
pthread_cond_t main_cond = PTHREAD_COND_INITIALIZER;
struct suxthr {
	struct sux sx;
	uint8_t th;
};
#if bench
double ipc;
struct timeval str[THREADS], en[THREADS];
#endif

void *run(void *args) {
	struct suxthr *thr = (void *)args;
	struct sux *cpu = &thr->sx;
	uint8_t thread = thr->th;
	uint64_t address = 0;
	uint64_t tmpaddr = 0;
	uint8_t prefix = 0;
	uint8_t opcode = 0;
	uint8_t end = 0;
	uint8_t stksize;
	uint64_t sum = 0;
	uint64_t value = 0;
	uint64_t reg = 0;
	uint64_t iclk = 0;
	uint64_t ins = 0;
	uint64_t sign = 0;
	uint8_t addrsize;
	uint8_t rs;
	uint8_t regsize;
	uint8_t tmp;
	char *s = malloc(2048);
#if !bench
	uint8_t lines = (6*thread)+2;
	uint8_t bcd[4];
	uint8_t idx = 3, iscol = 0;
	int x = 0, y = 0;
	uint8_t esc = 0;
#endif
	uint16_t tv = 0xFF50; /* Starting address of the Thread Vectors. */
#if bench
	gettimeofday(&str[thread], 0);
#endif
	while (!end) {
		address = 0;
		if (wai) {
			for (int8_t i = 56; i >= 0; i-=8) {
				if (i)
					addr[(cpu->stk_st[thread] << 16)+cpu->sp[thread]] = (uint64_t)cpu->pc[thread]-1 >> i;
				else
					addr[(cpu->stk_st[thread] << 16)+cpu->sp[thread]] = (uint64_t)cpu->pc[thread]-1 & 0xFF;
				cpu->sp[thread]--;
			}
			addr[(cpu->stk_st[thread] << 16)+cpu->sp[thread]] = (uint64_t)cpu->ps >> 8*thread;
			cpu->sp[thread]--;
			cpu->i[thread] = 1;
			setflag(cpu->i[thread], I);
			cpu->pc[thread] = (uint64_t)addr[0xFFA0]
					| (uint64_t)addr[0xFFA1] << 8
					| (uint64_t)addr[0xFFA2] << 16
					| (uint64_t)addr[0xFFA3] << 24
					| (uint64_t)addr[0xFFA4] << 32
					| (uint64_t)addr[0xFFA5] << 40
					| (uint64_t)addr[0xFFA6] << 48
					| (uint64_t)addr[0xFFA7] << 56;
			wai = 0;
			kbd_rdy &= (uint8_t)~(1 << thread);
		}
		prefix = addr[cpu->pc[thread]];
		if ((prefix & 0x07) == 0x07)
			cpu->pc[thread]++;
		else
			prefix = 0;
		opcode = addr[cpu->pc[thread]];


	#if debug && !bench
		if (lines > 24*(thread+1)) {
			lines = (24*thread)+2;
		}
		#if keypoll
		pthread_mutex_lock(&mutex);
		#endif
		/*mvwprintw(scr, lines, 0, "pc: $%08llx, a: $%016llx, b: $%016llx, x: $%016llx, y: $%016llx"
			", sp: $%04lx, ps: $%016llx, prefix: $%02x, opcode: $%02x, thread: %u, inst: "
			, cpu->pc[thread], cpu->a[thread], cpu->b[thread], cpu->x[thread], cpu->y[thread]
			, cpu->sp[thread], cpu->ps, prefix, opcode, thread);*/
		mvwprintw(scr, lines, 0, "pc: $%08llx, a: $%016llx, b: $%016llx, x: $%016llx, y: $%016llx"
			", sp: $%04lx, ps: $%016llx, prefix: $%02x, opcode: $%02x, inst: "
			, cpu->pc[thread], cpu->a[thread], cpu->b[thread], cpu->x[thread], cpu->y[thread]
			, cpu->sp[thread], cpu->ps, prefix, opcode);
		#if keypoll
		pthread_mutex_unlock(&mutex);
		#endif
	#endif

		addrsize = (prefix & 8) == 8;
		rs = (prefix & 0x30) >> 4;
		regsize = (1 << rs);
		address = cpu->pc[thread];
		cpu->pc[thread]++;
		iclk++;
		switch (optype[opcode]) {
			case IMPL:
				break;
			case IMM:
				switch (opcode) {
					case TXS:
						break;

					case PHB:
					case PHP:
					case PHA:
					case PHY:
					case PHX:
					case PLB:
					case PLP:
					case PLA:
					case PLY:
					case PLX:
					case STT:
					case LSL:
					case LSR:
					case ROL:
					case ROR:
					case ASR:
					case ENT:
						address = cpu->pc[thread];
						cpu->pc[thread]+=1;
						break;
					default:
						address = cpu->pc[thread];
						cpu->pc[thread]+=regsize;
						break;
				}
				break;
			case ZM:
			case ZMX:
			case ZMY:
				address = addr[cpu->pc[thread]];
				if (addrsize) {
					address |= addr[cpu->pc[thread]+1] << 8;
					address |= addr[cpu->pc[thread]+2] << 16;
					address |= addr[cpu->pc[thread]+3] << 24;
					cpu->pc[thread]+=4;
				} else {
					cpu->pc[thread]+=1;
				}
				tmpaddr = address;
				switch (optype[opcode]) {
					case ZMX:
						address += cpu->x[thread];
						iclk++;
						break;
					case ZMY:
						address += cpu->y[thread];
						iclk++;
						break;
				}
				iclk++;
				break;
			case IND:
			case INDX:
			case INDY:
				address = addr[cpu->pc[thread]];
				if (addrsize) {
					address |= addr[cpu->pc[thread]+1] << 8;
					address |= addr[cpu->pc[thread]+2] << 16;
					address |= addr[cpu->pc[thread]+3] << 24;
					cpu->pc[thread]+=4;
				} else {
					cpu->pc[thread]+=1;
				}
				iclk++;
				tmpaddr = address;
				if (optype[opcode] == INDX) {
					address += cpu->x[thread];
					iclk++;
				}
				value =	addr[address];
				value |= addr[address+1] << 8;
				value |= addr[address+2] << 16;
				value |= addr[address+3] << 24;
				value |= (uint64_t)addr[address+4] << 32;
				value |= (uint64_t)addr[address+5] << 40;
				value |= (uint64_t)addr[address+6] << 48;
				value |= (uint64_t)addr[address+7] << 56;
				iclk++;
				if (optype[opcode] == INDY) {
					value += cpu->y[thread];
					iclk++;
				}
				address = value;
				value = 0;
				iclk++;
				break;
			case ABS:
				address = addr[cpu->pc[thread]];
				address |= addr[cpu->pc[thread]+1] << 8;
				if (addrsize) {
					address |= addr[cpu->pc[thread]+2] << 16;
					address |= addr[cpu->pc[thread]+3] << 24;
					address |= (uint64_t)addr[cpu->pc[thread]+4] << 32;
					address |= (uint64_t)addr[cpu->pc[thread]+5] << 40;
					address |= (uint64_t)addr[cpu->pc[thread]+6] << 48;
					address |= (uint64_t)addr[cpu->pc[thread]+7] << 56;
					cpu->pc[thread]+=8;
					iclk++;
				} else {
					cpu->pc[thread]+=2;
				}
				iclk++;
				break;

		}
		value = addr[address];
		if (regsize >= 2) {
			value |= addr[address+1] << 8;
			if (regsize >= 4) {
				value |= addr[address+2] << 16;
				value |= addr[address+3] << 24;

				if (regsize >= 8) {
					value |= (uint64_t)addr[address+4] << 32;
					value |= (uint64_t)addr[address+5] << 40;
					value |= (uint64_t)addr[address+6] << 48;
					value |= (uint64_t)addr[address+7] << 56;
				}
			}
		}
	#if debug && !bench
		int row, col;
		#if keypoll
		pthread_mutex_lock(&mutex);
		#endif
		getyx(scr,row, col);
		char postfix[3];
		char op[4];
		op[0] = opname[opcode][0];
		op[1] = opname[opcode][1];
		op[2] = opname[opcode][2];
		op[3] = '\0';
		if (regsize == 1) {
			postfix[0] = '\0';
			postfix[1] = '\0';
			postfix[2] = '\0';
		} else {
			postfix[0] = '.';
			if (regsize == 2)
				postfix[1] = 'W';
			else if (regsize == 4)
				postfix[1] = 'D';
			else if (regsize == 8)
				postfix[1] = 'Q';
			postfix[2] = '\0';
		}
		switch (optype[opcode]) {
			case IMPL:
				mvwprintw(scr, lines, col, "%s                      \r" , opname[opcode]);
				break;
			case IMM:
				if (regsize == 1) {
					mvwprintw(scr, lines, col, "%s #$%02x                \r" , op, value);
				} else if (regsize == 2) {
					mvwprintw(scr, lines, col, "%s%s #$%04x            \r" , op, postfix, value);
				} else if (regsize == 4) {
					mvwprintw(scr, lines, col, "%s%s #$%08x        \r" , op, postfix, value);
				} else if (regsize == 8) {
					mvwprintw(scr, lines, col, "%s%s #$%016llx\r" , op, postfix, value);
				}
				break;
			case ZM:
			case ZMX:
			case ZMY:
				if (optype[opcode] == ZMX)
					tmpaddr = address - cpu->x[thread];
				else if (optype[opcode] == ZMY)
					tmpaddr = address - cpu->y[thread];
				if (addrsize)
					mvwprintw(scr, lines, col, "%s%s $%08x%s        \r" , op, postfix, tmpaddr, (optype[opcode] == ZM) ? "" : ((optype[opcode] == ZMX) ? ", x" : ", y"));
				else
					mvwprintw(scr, lines, col, "%s%s $%02x%s        \r" , op, postfix, tmpaddr, (optype[opcode] == ZM) ? "" : ((optype[opcode] == ZMX) ? ", x" : ", y"));
				break;
			case IND:
			case INDX:
			case INDY:
				if (addrsize)
					mvwprintw(scr, lines, col, "%s%s ($%08x%s        \r" , op, postfix, tmpaddr, (optype[opcode] == IND) ? ")" : ((optype[opcode] == INDX) ? ", x)" : "), y"));
				else
					mvwprintw(scr, lines, col, "%s%s ($%02x%s        \r" , op, postfix, tmpaddr, (optype[opcode] == IND) ? ")" : ((optype[opcode] == INDX) ? ", x)" : "), y"));
				break;
			case ABS:
				tmpaddr = address;
				if (addrsize)
					mvwprintw(scr, lines, col, "%s%s $%016llx\r" , op, postfix, value);
				else
					mvwprintw(scr, lines, col, "%s%s $%04x       \r" , op, postfix, tmpaddr);
				break;

		}
		mvwprintw(scr, 29, 0, "address: $%016llx, scr_row: %02u, scr_col: %02u, scr_str: %02u, scr_end: %02u\r", address, addr[0], addr[1], addr[0x4B], addr[0x4C]);
		mvwprintw(scr, 32, 0, "bitabl: %02x %02x %02x %02x %02x %02x %02x %02x "
				      "%02x %02x %02x %02x %02x %02x %02x %02x\r"
				       , addr[0x3B], addr[0x3C], addr[0x3D], addr[0x3E], addr[0x3F], addr[0x40], addr[0x41], addr[0x42]
				       , addr[0x43], addr[0x44], addr[0x45], addr[0x46], addr[0x47], addr[0x48], addr[0x49], addr[0x4A]);
		mvwprintw(scr, 33, 0, "buffer:\r");
		uint8_t ln = 34;
		uint16_t tmpad = 0x2000;
		for (uint16_t i = 0; i < 20; i+=2) {
			mvwprintw(scr, ln++, 0, "%02x %02x %02x %02x %02x %02x %02x %02x %02x %02x %02x %02x %02x %02x %02x %02x "
						"%02x %02x %02x %02x %02x %02x %02x %02x %02x %02x %02x %02x %02x %02x %02x %02x, i: %02x\r"
						, addr[tmpad+(0x00)+(i*0x10)], addr[tmpad+(0x01)+(i*0x10)], addr[tmpad+(0x02)+(i*0x10)], addr[tmpad+(0x03)+(i*0x10)]
						, addr[tmpad+(0x04)+(i*0x10)], addr[tmpad+(0x05)+(i*0x10)], addr[tmpad+(0x06)+(i*0x10)], addr[tmpad+(0x07)+(i*0x10)]
						, addr[tmpad+(0x08)+(i*0x10)], addr[tmpad+(0x09)+(i*0x10)], addr[tmpad+(0x0A)+(i*0x10)], addr[tmpad+(0x0B)+(i*0x10)]
						, addr[tmpad+(0x0C)+(i*0x10)], addr[tmpad+(0x0D)+(i*0x10)], addr[tmpad+(0x0E)+(i*0x10)], addr[tmpad+(0x0F)+(i*0x10)]
						, addr[tmpad+(0x10)+(i*0x10)], addr[tmpad+(0x11)+(i*0x10)], addr[tmpad+(0x12)+(i*0x10)], addr[tmpad+(0x13)+(i*0x10)]
						, addr[tmpad+(0x14)+(i*0x10)], addr[tmpad+(0x15)+(i*0x10)], addr[tmpad+(0x16)+(i*0x10)], addr[tmpad+(0x17)+(i*0x10)]
						, addr[tmpad+(0x18)+(i*0x10)], addr[tmpad+(0x19)+(i*0x10)], addr[tmpad+(0x1A)+(i*0x10)], addr[tmpad+(0x1B)+(i*0x10)]
						, addr[tmpad+(0x1C)+(i*0x10)], addr[tmpad+(0x1D)+(i*0x10)], addr[tmpad+(0x1E)+(i*0x10)], addr[tmpad+(0x1F)+(i*0x10)]
						, i);

		}
		/*tmpad = 0x4000;
		mvwprintw(scr, ln++, 0, "cmd_buf:\r");
		for (uint16_t i = 0; i < 20; i+=2) {
			mvwprintw(scr, ln++, 0, "%02x %02x %02x %02x %02x %02x %02x %02x %02x %02x %02x %02x %02x %02x %02x %02x "
						"%02x %02x %02x %02x %02x %02x %02x %02x %02x %02x %02x %02x %02x %02x %02x %02x, i: %02x\r"
						, addr[tmpad+(0x00)+(i*0x10)], addr[tmpad+(0x01)+(i*0x10)], addr[tmpad+(0x02)+(i*0x10)], addr[tmpad+(0x03)+(i*0x10)]
						, addr[tmpad+(0x04)+(i*0x10)], addr[tmpad+(0x05)+(i*0x10)], addr[tmpad+(0x06)+(i*0x10)], addr[tmpad+(0x07)+(i*0x10)]
						, addr[tmpad+(0x08)+(i*0x10)], addr[tmpad+(0x09)+(i*0x10)], addr[tmpad+(0x0A)+(i*0x10)], addr[tmpad+(0x0B)+(i*0x10)]
						, addr[tmpad+(0x0C)+(i*0x10)], addr[tmpad+(0x0D)+(i*0x10)], addr[tmpad+(0x0E)+(i*0x10)], addr[tmpad+(0x0F)+(i*0x10)]
						, addr[tmpad+(0x10)+(i*0x10)], addr[tmpad+(0x11)+(i*0x10)], addr[tmpad+(0x12)+(i*0x10)], addr[tmpad+(0x13)+(i*0x10)]
						, addr[tmpad+(0x14)+(i*0x10)], addr[tmpad+(0x15)+(i*0x10)], addr[tmpad+(0x16)+(i*0x10)], addr[tmpad+(0x17)+(i*0x10)]
						, addr[tmpad+(0x18)+(i*0x10)], addr[tmpad+(0x19)+(i*0x10)], addr[tmpad+(0x1A)+(i*0x10)], addr[tmpad+(0x1B)+(i*0x10)]
						, addr[tmpad+(0x1C)+(i*0x10)], addr[tmpad+(0x1D)+(i*0x10)], addr[tmpad+(0x1E)+(i*0x10)], addr[tmpad+(0x1F)+(i*0x10)]
						, i);

		}*/
		wrefresh(scr);
		#if keypoll
		pthread_mutex_unlock(&mutex);
		#endif
		lines+=1;
	#endif
		switch(opcode) {
			case CPS: /* Clear Processor Status. */
				clearflags(0);
				clearflags(1);
				clearflags(2);
				clearflags(3);
				clearflags(4);
				clearflags(5);
				clearflags(6);
				clearflags(7);
				cpu->ps &= 0;
				break;
			case ADC:  /* ADC Immediate. */
			case AAB: /* Add Accumulator with carry by B register. */
			case 0x03: /* ADC Absolute. */
			case 0x05: /* ADC Zero Matrix. */
				if (opcode == AAB)
					value = cpu->b[thread];
				sum = cpu->a[thread]+value+cpu->c[thread];
				cpu->a[thread] = sum;
				cpu->z[thread] = (sum == 0);
				cpu->n[thread] = (sum >> 63);
				cpu->v[thread] = !((cpu->a[thread]^value) & 0x8000000000000000) && ((cpu->a[thread]^sum) & 0x8000000000000000);
				cpu->c[thread] = (sum < value);
				setflag(cpu->z[thread], Z);
				setflag(cpu->n[thread], N);
				setflag(cpu->v[thread], V);
				setflag(cpu->c[thread], C);
			setflag(cpu->i[thread], I);
				break;
			case PHB: /* PusH B register to stack. */
			case PHP: /* PusH Processor status to stack. */
			case PHA: /* PusH Accumulator to stack. */
			case PHY: /* PusH Y register to stack. */
			case PHX: /* PusH X register to stack. */
				tmp = value;
				if (opcode == PHB) {
					reg = cpu->b[thread];
				} else if (opcode == PHP) {
					reg = cpu->ps;
				} else if (opcode == PHA) {
					reg = cpu->a[thread];
				} else if (opcode == PHY) {
					reg = cpu->y[thread];
				} else if (opcode == PHX) {
					reg = cpu->x[thread];
				}

				if (tmp > 7)
					tmp = 7;
				for (int8_t i = tmp<<3; i >= 0; i-=8) {
					if (i)
						addr[(cpu->stk_st[thread] << 16)+cpu->sp[thread]] = reg >> i;
					else
						addr[(cpu->stk_st[thread] << 16)+cpu->sp[thread]] = reg & 0xFF;
					cpu->sp[thread]--;
				}
				break;
			case TAY: /* Transfer Accumulator to Y. */
			case TAX: /* Transfer Accumulator to Y. */
			case TYX: /* Transfer Y to X. */
			case TYA: /* Transfer Y to Accumulator. */
			case TXA: /* Transfer X to Accumulator. */
			case TXY: /* Transfer X to Y. */
			case TAB: /* Transfer Accumulator to B. */
			case TSX: /* Transfer Stack pointer to X. */
			case TBA: /* Transfer B to Accumulator. */
			case TXS: /* Transfer X to Stack pointer. */
				switch (opcode) {
					case TAY:
						cpu->y[thread] = cpu->a[thread];
						break;
                                        case TAX:
						cpu->x[thread] = cpu->a[thread];
						break;
                                        case TYX:
						cpu->x[thread] = cpu->y[thread];
						break;
                                        case TYA:
						cpu->a[thread] = cpu->y[thread];
						break;
                                        case TXA:
						cpu->a[thread] = cpu->x[thread];
						break;
					case TXY:
						cpu->y[thread] = cpu->x[thread];
						break;
                                        case TAB:
						cpu->b[thread] = cpu->a[thread];
						cpu->z[thread] = (cpu->b[thread] == 0);
						cpu->n[thread] = (cpu->b[thread] >> 63);
						break;
                                        case TSX:
						cpu->x[thread] = cpu->sp[thread] & 0xFFFF;
						cpu->x[thread] = cpu->stk_st[thread] << 16;
						break;
                                        case TBA:
						cpu->a[thread] = cpu->b[thread];
						break;
                                        case TXS:
						cpu->sp[thread] = cpu->x[thread];
						if (prefix == 0x17 && (value == thread+1 || value > 8)) {
							cpu->stk_st[thread] = value & 0xFF;
							cpu->stk_st[thread] += value << 16;
							cpu->pc[thread]+=2;
						}
						break;
				}
				switch (opcode) {
					case TYA:
					case TXA:
					case TBA:
						cpu->z[thread] = (cpu->a[thread] == 0);
						cpu->n[thread] = (cpu->a[thread] >> 63);
						break;
					case TAY:
					case TXY:
						cpu->z[thread] = (cpu->y[thread] == 0);
						cpu->n[thread] = (cpu->y[thread] >> 63);
						break;
					case TAX:
					case TYX:
						cpu->z[thread] = (cpu->x[thread] == 0);
						cpu->n[thread] = (cpu->x[thread] >> 63);
						break;
				}
				setflag(cpu->z[thread], Z);
				setflag(cpu->n[thread], N);
				break;
			case JMP: /* JMP Absolute. */
				cpu->pc[thread] = address;
				break;
			case SBC: /* SBC Immediate. */
			case SAB: /* Subtract Accumulator with carry by B register. */
			case 0x13: /* SBC Absolute. */
			case 0x15: /* SBC Zero Matrix. */
				if (opcode == SAB) {
					value = cpu->b[thread];
				}
				sum = cpu->a[thread]-value-!cpu->c[thread];
				cpu->z[thread] = (sum == 0);
				cpu->n[thread] = (sum >> 63);
				cpu->v[thread] = ((cpu->a[thread]^value) & 0x8000000000000000) && ((cpu->a[thread]^sum) & 0x8000000000000000);
				cpu->c[thread] = (sum > value);
				setflag(cpu->z[thread], Z);
				setflag(cpu->n[thread], N);
				setflag(cpu->v[thread], V);
				setflag(cpu->c[thread], C);
				cpu->a[thread] = sum;
				break;
			case PLB: /* PuLl B register from stack. */
			case PLP: /* PuLl Processor status from stack. */
			case PLA: /* PuLl Accumulator from stack. */
			case PLY: /* PuLl Y register from stack. */
			case PLX: /* PuLl X register from stack. */
				tmp = value;
				if (opcode == PLB)
					reg = cpu->b[thread];
				else if (opcode == PLP)
					reg = cpu->ps;
				else if (opcode == PLA)
					reg = cpu->a[thread];
				else if (opcode == PLY)
					reg = cpu->y[thread];
				else if (opcode == PLX)
					reg = cpu->x[thread];

				if (tmp > 7)
					tmp = 7;
				for (uint8_t i = 0; i < (tmp+1)<<3; i+=8) {
					cpu->sp[thread]++;
					if (i)
						reg += addr[(cpu->stk_st[thread] << 16)+cpu->sp[thread]] << i;
					else
						reg = addr[(cpu->stk_st[thread] << 16)+cpu->sp[thread]] & 0xFF;
				}
				if (opcode == PLB) {
					cpu->b[thread] = reg;
				} else if (opcode == PLP) {
					cpu->ps = reg;
				} else if (opcode == PLA) {
					cpu->a[thread] = reg;
				} else if (opcode == PLY) {
					cpu->y[thread] = reg;
				} else if (opcode == PLX) {
					cpu->x[thread] = reg;
				}
				break;
			case 0x34: /* JSR Indirect. */
			case 0x44: /* JSR Indexed Indirect. */
			case 0x54: /* JSR Indirect Indexed. */
			case JSR: /* Jump to SubRoutine. */
				if (addrsize)
					stksize = 24;
				else
					stksize = 0;
				for (int8_t i = stksize; i >= 0; i-=8) {
					if (i)
						addr[(cpu->stk_st[thread] << 16)+cpu->sp[thread]] = (uint64_t)cpu->pc[thread] >> i;
					else
						addr[(cpu->stk_st[thread] << 16)+cpu->sp[thread]] = (uint64_t)cpu->pc[thread] & 0xFF;
					cpu->sp[thread]--;
				}
				cpu->pc[thread] = address;
				break;
			case AND: /* AND Immediate. */
			case ABA: /* bitwise And with Accumulator, and B register. */
			case 0x23: /* AND Absolute. */
			case 0x25: /* AND Zero Matrix. */
				if (opcode == ABA) {
					value = cpu->b[thread];
				}
				cpu->a[thread] &= value;
				cpu->z[thread] = (value == 0);
				cpu->n[thread] = (value >> 63);
				setflag(cpu->z[thread], Z);
				setflag(cpu->n[thread], N);
				break;
			case STT: /* STart Thread. */
				cpu->crt |= value;
				for (uint8_t i = 0; i < 7; i++) {
					if ((value >> i) & 1) {
						address = (uint64_t)addr[tv+(i<<3)]
								| (uint64_t)addr[tv+1+(i<<3)] << 8
								| (uint64_t)addr[tv+2+(i<<3)] << 16
								| (uint64_t)addr[tv+3+(i<<3)] << 24
								| (uint64_t)addr[tv+4+(i<<3)] << 32
								| (uint64_t)addr[tv+5+(i<<3)] << 40
								| (uint64_t)addr[tv+6+(i<<3)] << 48
								| (uint64_t)addr[tv+7+(i<<3)] << 56;
						cpu->pc[i+1] = address;
					}
				}
				break;
			case BPO: /* BPO Absolute. */
			case 0x64: /* BPO Zero Matrix. */
				if (!cpu->n[thread])
					cpu->pc[thread] = address;
				break;
			case ORA: /* ORA Immediate. */
			case OAB: /* bitwise Or with Accumulator, and B register. */
			case 0x33: /* ORA Absolute. */
			case 0x35: /* ORA Zero Matrix. */
				if (opcode == OAB) {
					value = cpu->b[thread];
				}
				cpu->a[thread] |= value;
				cpu->z[thread] = (value == 0);
				cpu->n[thread] = (value >> 63);
				setflag(cpu->z[thread], Z);
				setflag(cpu->n[thread], N);
				break;
			case SEI: /* SEt Interrupt. */
				cpu->i[thread] = 1;
				setflag(cpu->i[thread], I);
				break;
			case BNG: /* BNG Absolute. */
			case 0x74: /* BNG Zero Matrix. */
				if (cpu->n[thread])
					cpu->pc[thread] = address;
				break;
			case XOR: /* XOR Immediate. */
			case XAB: /* bitwise Xor with Accumulator, and B register. */
			case 0x43: /* XOR Absolute. */
			case 0x45: /* XOR Zero Matrix. */
				if (opcode == XAB) {
					value = cpu->b[thread];
				}
				cpu->a[thread] ^= value;
				cpu->z[thread] = (value == 0);
				cpu->n[thread] = (value >> 63);
				setflag(cpu->z[thread], Z);
				setflag(cpu->n[thread], N);
				break;
			case CLI: /* CLear Interrupt. */
				cpu->i[thread] = 0;
				setflag(cpu->i[thread], I);
				break;
			case BCS: /* BCS Absolute. */
			case 0x84: /* BCS Zero Matrix. */
				if (cpu->c[thread])
					cpu->pc[thread] = address;
				break;
			case LSL: /* LSL Immediate. */
			case LLB: /* Logical shift Left accumulator by B. */
			case 0x53: /* LSL Absolute. */
			case 0x55: /* LSL Zero Matrix. */
				if (opcode == LLB) {
					value = cpu->b[thread];
				}
				sum = (value < 64) ? cpu->a[thread] << value : 0;
				cpu->z[thread] = (sum == 0);
				cpu->n[thread] = (sum >> 63);
				cpu->c[thread] = cpu->a[thread] >> 64-value;
				cpu->a[thread] = sum;
				setflag(cpu->z[thread], Z);
				setflag(cpu->n[thread], N);
				setflag(cpu->c[thread], C);
				break;
			case SEC: /* SEt Carry flag.*/
				cpu->c[thread] = 1;
				setflag(cpu->c[thread], C);
				break;
			case STA: /* STA Absolute. */
			case STY: /* STY Absolute. */
			case STX: /* STX Absolute. */
			case STB: /* STB Absolute. */
			case 0x49: /* STA Zero Matrix. */
			case 0x4A: /* STY Zero Matrix. */
			case 0x4B: /* STX Zero Matrix. */
			case 0x4E: /* STB Zero Matrix. */
			case 0x69: /* STA Zero Matrix, Indexed with X. */
			case 0x6A: /* STY Zero Matrix, Indexed with X. */
			case 0x6B: /* STX Zero Matrix, Indexed with Y. */
			case 0x6E: /* STB Zero Matrix, Indexed with X. */
			case 0x89: /* STA Zero Matrix, Indexed with Y. */
			case 0x8A: /* STY Indirect. */
			case 0x8B: /* STX Indirect. */
			case 0x8E: /* STB Zero Matrix, Indexed with Y. */
			case 0xA9: /* STA Indirect. */
			case 0xAA: /* STY Indexed Indirect. */
			case 0xAB: /* STX Indirect Indexed. */
			case 0xAE: /* STB Indirect. */
			case 0xC9: /* STA Indexed Indirect. */
			case 0xCE: /* STB Indexed Indirect. */
			case 0xE9: /* STA Indirect Indexed. */
			case 0xEE: /* STB Indirect Indexed. */
				switch (opcode) {
					case STB:
					case 0x4E:
					case 0x6E:
					case 0x8E:
					case 0xAE:
					case 0xCE:
					case 0xEE:
						value = cpu->b[thread];
						break;
					case STA:
					case 0x49:
					case 0x69:
					case 0x89:
					case 0xA9:
					case 0xC9:
					case 0xE9:
						value = cpu->a[thread];
						break;
					case STY:
					case 0x4A:
					case 0x6A:
					case 0x8A:
					case 0xAA:
						value = cpu->y[thread];
						break;

					case STX:
					case 0x4B:
					case 0x6B:
					case 0x8B:
					case 0xAB:
						value = cpu->x[thread];
						break;
				}
				addr[address] = value & 0xFF;
				if (address == STEP_ADDR) {
					step = addr[address];
				}
				#if IO
				if (address == TX_ADDR) {
					#if keypoll
					pthread_mutex_lock(&mutex);
					#endif
					if (esc) {
						switch(addr[address]) {
							case 'A':
								if (y > 0)
									y--;
								wmove(scr, y, x);
								esc = 0;
								break;
							case 'B':
								if (y < getmaxy(scr))
									y++;
								wmove(scr, y, x);
								esc = 0;
								break;
							case 'C':
								if (x < getmaxx(scr))
									x++;
								wmove(scr, y, x);
								esc = 0;
								break;
							case 'D':
								if (x > 0)
									x--;
								wmove(scr, y, x);
								esc = 0;
								break;
							case 'H':
								if (!bcd[2] && !bcd[3])
									y = 0;
								else
									y = ((bcd[3]*10) + bcd[2]);
								if (!bcd[0] && !bcd[1])
									x = 0;
								else
									x = ((bcd[1]*10) + bcd[0]);
								idx = 3;
								wmove(scr, y, x);
								bcd[0] = 0;
								bcd[1] = 0;
								bcd[2] = 0;
								bcd[3] = 0;
								esc = 0;
								break;
							case 'S':
								wscrl(scr, -1);
								esc = 0;
								break;
							case 'T':
								wscrl(scr, 1);
								esc = 0;
								break;
							case '0':
							case '1':
							case '2':
							case '3':
							case '4':
							case '5':
							case '6':
							case '7':
							case '8':
							case '9':
								bcd[idx--] = (addr[address] - '0');
								break;
							default:
								iscol = (addr[address] == ';');
								break;
						}
					} else {
						if (addr[address] == 0xC) {
							x=0,y=0;
							wclear(scr);
							wmove(scr, y, x);
						} else if (addr[address] == CURSES_BACKSPACE || addr[address] == '\b') {
							if (x > 0) {
								x--;
								wmove(scr, y, x);
							}
							wdelch(scr);
							wrefresh(scr);
						} else if (addr[address] == '\033') {
							esc = 1;
						} else {
							wmove(scr, y, x);
							waddch(scr, addr[address]);
							wrefresh(scr);
							if (addr[address] == '\n') {
								x = 0;
								y+=1;
							} else {
								x+=1;
							}
						}
					}
					#if keypoll
					pthread_mutex_unlock(&mutex);
					#endif
				}
				#else
				#if !bench
				if (address == TX_ADDR) {
					if (esc) {
						switch(addr[address]) {
							case 'A':
								if (y > 0)
									y--;
								esc = 0;
								break;
							case 'B':
								if (y < getmaxy(scr))
									y++;
								esc = 0;
								break;
							case 'C':
								if (x < getmaxx(scr))
									x++;
								esc = 0;
								break;
							case 'D':
								if (x > 0)
									x--;
								esc = 0;
								break;
							case 'H':
								if (!bcd[2] && !bcd[3])
									y = 0;
								else
									y = ((bcd[3]*10) + bcd[2]);
								if (!bcd[0] && !bcd[1])
									x = 0;
								else
									x = ((bcd[1]*10) + bcd[0]);
								mvwprintw(scr, 30, 0, "x: %i, y: %i  ", x, y);
								mvwprintw(scr, 31, 0, "bcd[3-2]: {%u, %u}, bcd[1-0]: {%u, %u}", bcd[3], bcd[2], bcd[1], bcd[0]);
								wrefresh(scr);
								idx = 3;
								bcd[0] = 0;
								bcd[1] = 0;
								bcd[2] = 0;
								bcd[3] = 0;
								esc = 0;
								break;
							case 'S':
							case 'T':
								esc = 0;
								break;
							case '0':
							case '1':
							case '2':
							case '3':
							case '4':
							case '5':
							case '6':
							case '7':
							case '8':
							case '9':
								bcd[idx--] = (addr[address] - '0');
								break;
							default:
								iscol = (addr[address] == ';');
								break;
						}
					} else {
						if (addr[address] == CURSES_BACKSPACE || addr[address] == '\b') {
							if (x > 0) {
								x--;
							}
						} else if (addr[address] == '\033') {
							esc = 1;
						} else {
							if (addr[address] == '\n') {
								x = 0;
								y+=1;
							} else {
								x+=1;
							}
						}
					}
				}
				#endif
				#endif
				if (regsize >= 2) {
					addr[address+1] = value >> 8;
					if (regsize >= 4) {
						addr[address+2] = value >> 16;
						addr[address+3] = value >> 24;
						if (regsize >= 8) {
							addr[address+4] = value >> 32;
							addr[address+5] = value >> 40;
							addr[address+6] = value >> 48;
							addr[address+7] = value >> 56;
						}
					}
				}
				break;
			case BCC: /* BCC Absolute. */
			case 0x94: /* BCC Zero Matrix. */
				if (!cpu->c[thread])
					cpu->pc[thread] = address;
				break;
			case LSR: /* LSR Immediate. */
			case LRB: /* Logical shift Right accumulator by B. */
			case 0x63: /* LSR Absolute. */
			case 0x65: /* LSR Zero Matrix. */
				if (opcode == LRB) {
					value = cpu->b[thread];
				}
				sum = (value < 64) ? cpu->a[thread] >> value : 0;
				cpu->z[thread] = (sum == 0);
				cpu->n[thread] = (sum >> 63);
				cpu->c[thread] = cpu->a[thread] & 1;
				cpu->a[thread] = sum;
				setflag(cpu->z[thread], Z);
				setflag(cpu->n[thread], N);
				setflag(cpu->c[thread], C);
				break;
			case ASR: /* ASR Immediate. */
			case ARB: /* Arithmetic shift Right accumulator by B. */
			case 0xE3: /* ASR Absolute. */
			case 0xE5: /* ASR Zero Matrix. */
				if (opcode == ARB)
					value = cpu->b[thread];
				sign = cpu->a[thread] & 0x8000000000000000;
				sum = (value < 64) ? (cpu->a[thread] >> value) | sign : 0;
				cpu->z[thread] = (sum == 0);
				cpu->n[thread] = (sum >> 63);
				cpu->c[thread] = cpu->a[thread] & 1;
				cpu->a[thread] = sum;
				setflag(cpu->z[thread], Z);
				setflag(cpu->n[thread], N);
				setflag(cpu->c[thread], C);
				break;
			case CLC: /* CLear Carry flag. */
				cpu->c[thread] = 0;
				setflag(cpu->c[thread], C);
				break;
			case LDB: /* LDB Immediate. */
			case LDA: /* LDA Immediate. */
			case LDY: /* LDY Immediate. */
			case LDX: /* LDX Immediate. */
			case 0x1E: /* LDB Absolute. */
			case 0x19: /* LDA Absolute. */
			case 0x1A: /* LDY Absolute. */
			case 0x1B: /* LDX Absolute. */
			case 0x3E: /* LDB Zero Matrix. */
			case 0x39: /* LDA Zero Matrix. */
			case 0x3A: /* LDY Zero Matrix. */
			case 0x3B: /* LDX Zero Matrix. */
			case 0x5E: /* LDB Zero Matrix, Indexed with X. */
			case 0x59: /* LDA Zero Matrix, Indexed with X. */
			case 0x5A: /* LDY Zero Matrix, Indexed with X. */
			case 0x5B: /* LDX Zero Matrix, Indexed with Y. */
			case 0x7E: /* LDB Zero Matrix, Indexed with Y. */
			case 0x79: /* LDA Zero Matrix, Indexed with Y. */
			case 0x7A: /* LDY Indirect. */
			case 0x7B: /* LDX Indirect. */
			case 0x9E: /* LDB Indirect. */
			case 0x99: /* LDA Indirect. */
			case 0x9A: /* LDY Indexed Indirect. */
			case 0x9B: /* LDX Indirect Indexed. */
			case 0xBE: /* LDB Indexed Indirect. */
			case 0xB9: /* LDA Indexed Indirect. */
			case 0xDE: /* LDB Indirect Indexed. */
			case 0xD9: /* LDA Indirect Indexed. */
				if (address == CTRL_ADDR) {
					kbd_rdy = 1;
					pthread_mutex_lock(&main_mutex);
					pthread_cond_signal(&main_cond);
					pthread_mutex_unlock(&main_mutex);
					#if !keypoll
					pthread_mutex_lock(&mutex);
					pthread_cond_wait(&cond, &mutex);
					pthread_mutex_unlock(&mutex);
					#endif
					kbd_rdy = 0;
				}
				value = addr[address];
				if (regsize >= 2) {
					value += addr[address+1] << 8;
					if (regsize >= 4) {
						value += addr[address+2] << 16;
						value += addr[address+3] << 24;
						if (regsize >= 8) {
							value += (uint64_t)addr[address+4] << 32;
							value += (uint64_t)addr[address+5] << 40;
							value += (uint64_t)addr[address+6] << 48;
							value += (uint64_t)addr[address+7] << 56;
						}
					}
				}
				switch (opcode) {
					case LDB:
					case 0x1E:
					case 0x3E:
					case 0x5E:
					case 0x7E:
					case 0x9E:
					case 0xBE:
					case 0xDE:
						cpu->b[thread] = value;
						break;
					case LDA:
					case 0x19:
					case 0x39:
					case 0x59:
					case 0x79:
					case 0x99:
					case 0xB9:
					case 0xD9:
						cpu->a[thread] = value;
						break;
					case LDY:
					case 0x1A:
					case 0x3A:
					case 0x5A:
					case 0x7A:
					case 0x9A:
						cpu->y[thread] = value;
						break;

					case LDX:
					case 0x1B:
					case 0x3B:
					case 0x5B:
					case 0x7B:
					case 0x9B:
						cpu->x[thread] = value;
						break;
				}
				cpu->z[thread] = (value == 0);
				cpu->n[thread] = (value >> 63);
				setflag(cpu->z[thread], Z);
				setflag(cpu->n[thread], N);
				break;
			case BEQ: /* BEQ Absolute. */
			case 0xA4: /* BEQ Zero Matrix. */
				if (cpu->z[thread])
					cpu->pc[thread] = address;
				break;
			case ROL: /* ROL Immediate. */
			case RLB: /* Rotate Left accumulator by B. */
			case 0x73: /* ROL Absolute. */
			case 0x75: /* ROL Zero Matrix. */
				if (opcode == RLB) {
					value = cpu->b[thread];
				}
				sum = cpu->a[thread] << value;
				sum |= cpu->c[thread];
				cpu->z[thread] = (sum == 0);
				cpu->n[thread] = (sum >> 63);
				cpu->c[thread] = cpu->a[thread] >> (uint64_t)64-value;
				cpu->a[thread] = sum;
				setflag(cpu->z[thread], Z);
				setflag(cpu->n[thread], N);
				setflag(cpu->c[thread], C);
				break;
			case SSP: /* Set Stack Protection flag. */
				cpu->s[thread] = 1;
				setflag(cpu->s[thread], S);
				break;
			case BNE: /* BNE Absolute. */
			case 0xB4: /* BNE Zero Matrix. */
				if (!cpu->z[thread])
					cpu->pc[thread] = address;
				break;
			case ROR: /* ROR Immediate. */
			case RRB: /* Rotate Right accumulator by B. */
			case 0x83: /* ROR Absolute. */
			case 0x85: /* ROR Zero Matrix. */
				if (opcode == RRB) {
					value = cpu->b[thread];
				}
				sum = cpu->a[thread] >> value;
				sum |= (uint64_t)cpu->c[thread] << (uint64_t)64-value;
				cpu->z[thread] = (sum == 0);
				cpu->n[thread] = (sum >> 63);
				cpu->c[thread] = cpu->a[thread] & 1;
				cpu->a[thread] = sum;
				setflag(cpu->z[thread], Z);
				setflag(cpu->n[thread], N);
				setflag(cpu->c[thread], C);
				break;
			case CSP: /* Clear Stack Protection flag. */
				cpu->s[thread] = 0;
				setflag(cpu->s[thread], S);
				break;
			case BVS: /* BVS Absolute. */
			case 0xC4: /* BVS Zero Matrix. */
				if (cpu->v[thread])
					cpu->pc[thread] = address;
				break;
			case MUL: /* MUL Immediate. */
			case MAB: /* Multiply Accumulator by B. */
			case 0x93: /* MUL Absolute. */
			case 0x95: /* MUL Zero Matrix. */
				if (opcode == MAB) {
					value = cpu->b[thread];
				}
				sum = cpu->a[thread]*value+cpu->c[thread];
				cpu->a[thread] = sum;
				cpu->z[thread] = (sum == 0);
				cpu->n[thread] = (sum >> 63);
				cpu->v[thread] = !((cpu->a[thread]^value) & 0x8000000000000000) && ((cpu->a[thread]^sum) & 0x8000000000000000);
				cpu->c[thread] = (!((cpu->a[thread]^sum) && (cpu->a[thread]^value)) && (cpu->a[thread] >= ((uint64_t)1 << 32) && value >= ((uint64_t)1 << 32)));
				setflag(cpu->z[thread], Z);
				setflag(cpu->n[thread], N);
				setflag(cpu->v[thread], V);
				setflag(cpu->c[thread], C);
				break;
			case SEV: /* SEt oVerflow flag. */
				cpu->v[thread] = 1;
				setflag(cpu->v[thread], V);
				break;
			case BVC: /* BVC Absolute. */
			case 0xD4: /* BVC Zero Matrix. */
				if (!cpu->v[thread])
					cpu->pc[thread] = address;
				break;
			case DIV: /* DIV Immediate. */
			case DAB: /* Divide Accumulator by B. */
			case 0xA3: /* DIV Absolute. */
			case 0xA5: /* DIV Zero Matrix. */
				if (opcode != DAB) {
					cpu->b[thread] = cpu->a[thread] % value;
				} else {
					value = cpu->b[thread];
					cpu->x[thread] = cpu->a[thread] % value;
				}
				sum = cpu->a[thread]/value;
				cpu->a[thread] = sum;
				cpu->z[thread] = (sum == 0);
				cpu->n[thread] = (sum >> 63);
				setflag(cpu->z[thread], Z);
				setflag(cpu->n[thread], N);
				break;
			case CLV: /* CLear oVerflow flag. */
				cpu->v[thread] = 0;
				setflag(cpu->v[thread], V);
				break;
			case RTS: /* ReTurn from Subroutine. */
				if (addrsize)
					stksize = 32;
				else
					stksize = 8;
				for (uint8_t i = 0; i < stksize; i+=8) {
					cpu->sp[thread]++;
					if (i)
						cpu->pc[thread] += addr[(cpu->stk_st[thread] << 16)+cpu->sp[thread]] << i;
					else
						cpu->pc[thread] = addr[(cpu->stk_st[thread] << 16)+cpu->sp[thread]];
				}
				break;
			case CPB: /* CPB Immediate. */
			case CMP: /* CMP Immediate. */
			case CAB: /* Compare Accumulator, and B. */
			case CPY: /* CPY Immediate. */
			case CPX: /* CPX Immediate. */
			case 0xCA: /* CPY Absolute. */
			case 0xCB: /* CPX Absolute. */
			case 0xB3: /* CMP Absolute. */
			case 0x36: /* CPB Absolute. */
			case 0xDA: /* CPY Zero Matrix. */
			case 0xDB: /* CPX Zero Matrix. */
			case 0xB5: /* CMP Zero Matrix. */
			case 0x46: /* CPB Zero Matrix. */
			case 0xF1: /* CMP Indirect. */
			case 0xEA: /* CPY Indirect. */
			case 0xEB: /* CPX Indirect. */
			case 0x56: /* CPB Indirect. */
			case 0xF3: /* CMP Indexed Indirect. */
			case 0xFA: /* CPY Indexed Indirect. */
			case 0x66: /* CPB Indexed Indirect. */
			case 0xF5: /* CMP Indirect Indexed. */
			case 0xFB: /* CPX Indirect Indexed. */
			case 0x76: /* CPB Indirect Indexed. */
				if (opcode == CAB) {
					value = cpu->b[thread];
				}
				switch (opcode) {
					case CPB:
					case 0x36:
					case 0x46:
					case 0x56:
					case 0x66:
					case 0x76:
						reg = cpu->b[thread];
						break;
					case CMP:
					case CAB:
					case 0xB3:
					case 0xB5:
					case 0xF1:
					case 0xF3:
					case 0xF5:
						reg = cpu->a[thread];
						break;
					case CPY:
					case 0xCA:
					case 0xDA:
					case 0xEA:
					case 0xFA:
						reg = cpu->y[thread];
						break;

					case LDX:
					case 0xCB:
					case 0xDB:
					case 0xEB:
					case 0xFB:
						reg = cpu->x[thread];
						break;
				}
				sum = reg-value;
				cpu->n[thread] = (sum & 0x8000000000000000);
				cpu->v[thread] = ((reg^value) & 0x8000000000000000) && ((reg^sum) & 0x8000000000000000);
				cpu->z[thread] = (sum == 0);
				cpu->c[thread] = (reg >= value);
				setflag(cpu->n[thread], N);
				setflag(cpu->v[thread], V);
				setflag(cpu->z[thread], Z);
				setflag(cpu->c[thread], C);
				break;
			case ENT: /* ENd Thread. */
				cpu->crt &= ~value;
				for (uint8_t i = 0; i < 7; i++)
					if ((value >> i) & 1)
						cpu->pc[i+1] = cpu->pc[0]+(i+1);
				break;
			case RTI: /* ReTurn from Interrupt routine. */
				cpu->sp[thread]++;
				cpu->ps = addr[(cpu->stk_st[thread] << 16)+cpu->sp[thread]] << 8*thread;
				for (uint8_t i = 0; i < 64; i+=8) {
					cpu->sp[thread]++;
					if (i)
						cpu->pc[thread] += addr[(cpu->stk_st[thread] << 16)+cpu->sp[thread]] << i;
					else
						cpu->pc[thread] = addr[(cpu->stk_st[thread] << 16)+cpu->sp[thread]];
				}
				break;
			case INC: /* INC Accumulator. */
			case IAB:
			case INY:
			case INX:
				if (opcode == INC || opcode == IAB) {
					cpu->a[thread]+=1;
					if (opcode == IAB)
						cpu->b[thread]+=1;
					cpu->z[thread] = (cpu->a[thread] == 0);
					cpu->n[thread] = (cpu->a[thread] >> 63);
				}
				else if (opcode == INY) {
					cpu->y[thread]+=1;
					cpu->z[thread] = (cpu->y[thread] == 0);
					cpu->n[thread] = (cpu->y[thread] >> 63);
				}
				else if (opcode == INX) {
					cpu->x[thread]+=1;
					cpu->z[thread] = (cpu->x[thread] == 0);
					cpu->n[thread] = (cpu->x[thread] >> 63);
				}
				setflag(cpu->z[thread], Z);
				setflag(cpu->n[thread], N);
				break;
			case 0x04: /* JMP Indirect. */
			case 0x14: /* JMP Indexed Indirect. */
			case 0x24: /* JMP Indirect Indexed. */
			case 0xD0: /* JMP Zero Matrix. */
				cpu->pc[thread] = address;
				break;

			case DEC: /* DEC Accumulator. */
			case DBA:
			case DEY:
			case DEX:
				if (opcode == DEC || opcode == DBA) {
					cpu->a[thread]-=1;
					if (opcode == DBA)
						cpu->b[thread]-=1;
					cpu->z[thread] = (cpu->a[thread] == 0);
					cpu->n[thread] = (cpu->a[thread] >> 63);
				}
				else if (opcode == DEY) {
					cpu->y[thread]-=1;
					cpu->z[thread] = (cpu->y[thread] == 0);
					cpu->n[thread] = (cpu->y[thread] >> 63);
				}
				else if (opcode == DEX) {
					cpu->x[thread]-=1;
					cpu->z[thread] = (cpu->x[thread] == 0);
					cpu->n[thread] = (cpu->x[thread] >> 63);
				}
				setflag(cpu->z[thread], Z);
				setflag(cpu->n[thread], N);
				break;
			case JSL: /* Jump to Subroutine Long. */
				if (addrsize)
					stksize = 56;
				else
					stksize = 8;
				for (int8_t i = stksize; i >= 0; i-=8) {
					if (i)
						addr[(cpu->stk_st[thread] << 16)+cpu->sp[thread]] = (uint64_t)cpu->pc[thread] >> i;
					else
						addr[(cpu->stk_st[thread] << 16)+cpu->sp[thread]] = (uint64_t)cpu->pc[thread] & 0xFF;
					cpu->sp[thread]--;
				}
				cpu->pc[thread] = address;
				break;
			case 0xC3: /* INC Absolute. */
			case 0xC5: /* INC Zero Matrix. */
				addr[address]++;
				cpu->z[thread] = (addr[address] == 0);
				cpu->n[thread] = (addr[address] >> 7);
				setflag(cpu->z[thread], Z);
				setflag(cpu->n[thread], N);
				break;
			case NOP: /* No OPeration. */
				break;
			case RTL: /* ReTurn from subroutine Long. */
				if (addrsize)
					stksize = 64;
				else
					stksize = 16;
				for (uint8_t i = 0; i < stksize; i+=8) {
					cpu->sp[thread]++;
					if (i)
						cpu->pc[thread] += addr[(cpu->stk_st[thread] << 16)+cpu->sp[thread]] << i;
					else
						cpu->pc[thread] = addr[(cpu->stk_st[thread] << 16)+cpu->sp[thread]];
				}
				break;
			case 0xD3: /* DEC Absolute. */
			case 0xD5: /* DEC Zero Matrix. */
				addr[address]--;
				cpu->z[thread] = (addr[address] == 0);
				cpu->n[thread] = (addr[address] >> 7);
				setflag(cpu->z[thread], Z);
				setflag(cpu->n[thread], N);
				break;
			case BRK: /* BReaK. */
				for (int8_t i = stksize; i >= 0; i-=8) {
					if (i)
						addr[(cpu->stk_st[thread] << 16)+cpu->sp[thread]] = (uint64_t)cpu->pc[thread] >> i;
					else
						addr[(cpu->stk_st[thread] << 16)+cpu->sp[thread]] = (uint64_t)cpu->pc[thread] & 0xFF;
					cpu->sp[thread]--;
				}
				addr[(cpu->stk_st[thread] << 16)+cpu->sp[thread]] = (uint64_t)cpu->ps >> thread << 3;
				cpu->sp[thread]--;
				cpu->i[thread] = 1;
				setflag(cpu->i[thread], I);
				cpu->pc[thread] = (uint64_t)addr[0xFFE0]
						| (uint64_t)addr[0xFFE1] << 8
						| (uint64_t)addr[0xFFE2] << 16
						| (uint64_t)addr[0xFFE3] << 24
						| (uint64_t)addr[0xFFE4] << 32
						| (uint64_t)addr[0xFFE5] << 40
						| (uint64_t)addr[0xFFE6] << 48
						| (uint64_t)addr[0xFFE7] << 56;
				break;
			case WAI: /* WAit for Interrupt. */
				wai = 1;
				pthread_mutex_lock(&main_mutex);
				pthread_cond_signal(&main_cond);
				pthread_mutex_unlock(&main_mutex);
				pthread_mutex_lock(&mutex);
				pthread_cond_wait(&cond, &mutex);
				pthread_mutex_unlock(&mutex);
				break;
			default:
				break;
		}
		ins++;
		if (step) {
			pthread_mutex_lock(&main_mutex);
			pthread_cond_signal(&main_cond);
			pthread_mutex_unlock(&main_mutex);
			pthread_mutex_lock(&mutex);
			pthread_cond_wait(&cond, &mutex);
			pthread_mutex_unlock(&mutex);
		}
		if (!addr[CTRL_ADDR])
			kbd_ln = 0;
		else
			kbd_ln = 1;
		/*if (kbd_ln)
			usleep(16666);
			usleep(500000);*/
	#if debug && !bench
		#if keypoll
		pthread_mutex_lock(&mutex);
		#endif
		/*mvwprintw(scr, getmaxy(scr)-lines, 0, "Operand: $%llx"
				", $%04llx: $%02x, $%04llx: $%02x"
				", $1000: $%02x, $1001: $%02x    "
				, value
				, RX_ADDR, addr[RX_ADDR], TX_ADDR, addr[TX_ADDR]
				, addr[0x1000], addr[0x1001]);*/
		mvwprintw(scr, (6*thread)+1, 0, "Instructions executed: %llu, Clock cycles: %llu\r", ins, iclk);
		wrefresh(scr);
		#if keypoll
		pthread_mutex_unlock(&mutex);
		#endif
	#endif
	#if bench
		if (ins >= BENCH_INST) {
			end = 1;
			pthread_mutex_lock(&main_mutex);
			threads_done++;
			inst[thread] = ins;
			clk[thread] = iclk;
			pthread_cond_signal(&main_cond);
			pthread_mutex_unlock(&main_mutex);
			gettimeofday(&en[thread], 0);
		}
	#endif
	}
	free(s);
}

int main(int argc, char **argv) {
	struct suxthr thr[THREADS];
	char *tmp = malloc(2048);
	ibcount = 0;
	addr = malloc(0x04000000);
	inss = 0;
	int v = 0;

	if (argc != 2) {
		if (asmmon("stdin") == 2)
			return 0;
	} else {
		if (asmmon(argv[1]) == 2)
			return 0;
	}
	sprintf(tmp, "\033[2J\033[H");
	fwrite(tmp, sizeof(char), strlen(tmp), stdout);
	fflush(stdout);
#if !bench
	if(!scr)
		scr = initscr();
	nodelay(stdscr, 0);
	crmode();
	noecho();
	nl();
	curs_set(1);
	werase(scr);
	scrollok(scr, 1);
	wrefresh(scr);
	start_color();
	use_default_colors();
	init_pair(1, COLOR_WHITE, -1);
	attron(COLOR_PAIR(1) | A_BOLD);
	wmove(scr, 0, 0);
#endif
	for (int i = 0; i < THREADS; i++) {
		thr[i].sx.sp[i] = 0xFFFF;
		thr[i].sx.stk_st[i] = i+1;
		if (i) {
			thr[i].sx.a[i] = 0;
			thr[i].sx.b[i] = 0;
			thr[i].sx.x[i] = 0;
			thr[i].sx.y[i] = 0;
			thr[i].sx.pc[i] = (uint64_t)addr[0xFF50+(8*(i-1))]
					| (uint64_t)addr[0xFF51+(8*(i-1))] << 8
					| (uint64_t)addr[0xFF52+(8*(i-1))] << 16
					| (uint64_t)addr[0xFF53+(8*(i-1))] << 24
					| (uint64_t)addr[0xFF54+(8*(i-1))] << 32
					| (uint64_t)addr[0xFF55+(8*(i-1))] << 40
					| (uint64_t)addr[0xFF56+(8*(i-1))] << 48
					| (uint64_t)addr[0xFF57+(8*(i-1))] << 56;
		} else {
			thr[i].sx.a[i] = 0;
			thr[i].sx.b[i] = 0;
			thr[i].sx.x[i] = 0;
			thr[i].sx.y[i] = 0;
			thr[i].sx.pc[i] = (uint64_t)addr[0xFFC0]
					| (uint64_t)addr[0xFFC1] << 8
					| (uint64_t)addr[0xFFC2] << 16
					| (uint64_t)addr[0xFFC3] << 24
					| (uint64_t)addr[0xFFC4] << 32
					| (uint64_t)addr[0xFFC5] << 40
					| (uint64_t)addr[0xFFC6] << 48
					| (uint64_t)addr[0xFFC7] << 56;
		}
		thr[i].th = i;
	}
	pthread_t therads[THREADS];
	int result;
	for (int i = 0; i < THREADS; i++) {
		inst[i] = 0;
	}
	for (int i = 0; i < THREADS; i++) {
		result = pthread_create(&therads[i], NULL, run, &thr[i]);
		assert(!result);
	}
	int c = 0;
	uint8_t step_key = 0;
#if !bench
	werase(scr);
#endif
	while (threads_done < THREADS) {
		#if !bench
		int x, y, i = 0;
		if ((step_key && step && !kbd_rdy) || !step || kbd_rdy) {
			if ((c != EOF && c !=-1)) {
				pthread_mutex_lock(&main_mutex);
				curs_set(0);
				pthread_cond_wait(&main_cond, &main_mutex);
				pthread_mutex_unlock(&main_mutex);
				curs_set(1);
				c = 0;
				step_key = 0;
				addr[CTRL_ADDR] = 0;
				kbd_ln = 0;
			}
		}
		#if keypoll
		pthread_mutex_lock(&mutex);
		#endif
		getyx(scr, y, x);
		c = wgetch(scr);
		if (c == 19) {
			step = 1;
			if (kbd_rdy)
				c = wgetch(scr);
		}
		if (kbd_rdy) {
			switch (c) {
				case ERR:
					addr[CTRL_ADDR] = 0;
					wmove(scr, getmaxy(scr)-1, 0);
					wprintw(scr, "c: %i, x: %i, y: %i, i: %i.", c, x, y, i++);
					wmove(scr, y, x);
					wrefresh(scr);
					break;
				default:
					if (kbd_rdy && c < 0x100) {
						addr[RX_ADDR] = (uint8_t)c;
						addr[CTRL_ADDR] = 1;
					#if debug && !bench
						wmove(scr, getmaxy(scr)-1, 0);
						wprintw(scr, "c: %i, x: %i, y: %i, i: %i.", c, x, y, i++);
						wmove(scr, y, x);
						wrefresh(scr);
					#endif
					}
					if (c == '\n')
						kbd_ln = 1;
					#if !keypoll
					pthread_mutex_lock(&mutex);
					pthread_cond_signal(&cond);
					pthread_mutex_unlock(&mutex);
					#endif
					break;
			}
		} else {
			if ((c == 19 || c == 18) && step) {
				if (c == 18)
					step = 0;
				else if (c == 19)
					step_key = 1;
				#if !keypoll
				pthread_mutex_lock(&mutex);
				pthread_cond_signal(&cond);
				pthread_mutex_unlock(&mutex);
				#endif
			}
		}
		addr[STEP_ADDR] = step;
		#if keypoll
		pthread_mutex_unlock(&mutex);
		#endif
		#else
		pthread_mutex_lock(&main_mutex);
		pthread_cond_wait(&main_cond, &main_mutex);
		pthread_mutex_unlock(&main_mutex);
		#endif
	}
#if bench
	if (threads_done == THREADS) {
		double tm_sec, tm_usec, tm[THREADS], ttm;
		double clkspd;
		double mhz;
		double ips[THREADS];
		double ipst;
		for (int i = 0; i < THREADS; i++) {
			tm_sec = (en[i].tv_sec - str[i].tv_sec);
			tm_usec = (en[i].tv_usec-str[i].tv_usec);
			tm[i] = (tm_sec*1000000)+(tm_usec);
			ips[i] = inst[i]/tm[i];
			if (i) {
				inss += inst[i];
				ttm += tm[i];
				ipst += ips[i];
				tclk += clk[i];
			} else {
				inss = inst[i];
				ttm = tm[i];
				ipst = ips[i];
				tclk = clk[i];
			}
			clkspd = (tm[i]/1000000)*1000000/clk[i];
			mhz = 1000000.0/clkspd/1000000;
			sprintf(tmp, "Instructions executed for thread %i: %llu, Instructions per Second for thread %i in MIPS: %f, tm: %f\n", i, inst[i], i, ips[i], tm[i]/1000000);
			fwrite(tmp, sizeof(char), strlen(tmp), stdout);
		}
		clkspd = (ttm/1000000)*1000000/tclk;
		mhz = 1000000.0/clkspd/1000000;
		sprintf(tmp, "Total Instructions executed: %llu, Total Instructions per Second in MIPS: %f, Clock cycles: %llu, Clock Speed in MHz: %f, tm: %f\n", inss, ipst, tclk, mhz, ttm/1000000);
		fwrite(tmp, sizeof(char), strlen(tmp), stdout);
		fflush(stdout);
		free(tmp);
	}
#endif
	free(addr);
	return 0;
}