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

#define MAX(a, b)	((a > b) ? a : b)

#define STATUS_UPDATE_ZERO(r)		\
	(regs.status.zero = r == 0)
#define STATUS_UPDATE_NEGATIVE(r)	\
	(regs.status.negative = ((r & (1 << 7)) != 0))

#define MEMORY_MIRROR(addr)	\
	if (addr < 0x2000)	\
		addr &= 0x07FF;	\
	else if (addr < 0x4000)	\
		addr &= 0x2007;

#define PUSH(b)	\
	(memwrite(0x0100 + regs.sp--, b))
#define PULL()	\
	(peek(0x0100 + regs.sp++))

struct cpu_flags {
	uint8_t carry : 1;
	uint8_t zero : 1;
	uint8_t interrupt_disable : 1;
	uint8_t decimal_mode : 1;
	uint8_t brk : 1;
	uint8_t unused : 1;
	uint8_t overflow : 1;
	uint8_t negative : 1;
};

struct registers {
	uint8_t a, x, y, sp;
	struct cpu_flags status;
	uint16_t pc;
};
struct registers regs;

enum addressing_mode {
	AM_IMM,
	AM_ZP,
	AM_ZP_X,
	AM_ZP_Y,
	AM_REL,
	AM_ABS,
	AM_ABS_X,
	AM_ABS_Y,
	AM_IND,
	AM_IND_X,
	AM_IND_Y,
};

/* 64K address space, 16bit words */
uint8_t memory[0x16000];

/* example program taken from https://bugzmanov.github.io/nes_ebook/chapter_3_1.html */
uint8_t program[] = { 0xa9, 0xc0, 0xaa, 0xe8, 0x00 };

uint32_t cycles = 0;

static uint8_t
peek(uint16_t addr)
{
	MEMORY_MIRROR(addr);

	return memory[addr];
}

static uint16_t
peek16(uint16_t addr)
{
	MEMORY_MIRROR(addr);

	/* bytes are stored in little-endian (low then high) */
	return (uint16_t)memory[addr] | ((uint16_t)memory[addr + 1] << 8);
}

static void
memwrite(uint16_t addr, uint8_t byte)
{
	MEMORY_MIRROR(addr);

	memory[addr] = byte;
}

static void
memwrite16(uint16_t addr, uint16_t word)
{
	MEMORY_MIRROR(addr);

	/* bytes are stored in little-endian (low then high) */
	memory[addr] = word & 0xFF;
	memory[addr + 1] = (word & 0xFF00) >> 8;
}

static uint8_t
opcode_arg(enum addressing_mode mode)
{
	uint8_t arg, val;

	if (mode != AM_ABS && mode != AM_ABS_X && mode != AM_ABS_Y)
		arg = peek(regs.pc++);
	else
		arg = peek16(regs.pc), regs.pc += 2;

	switch (mode) {
	case AM_IMM:
	case AM_REL:
		val = arg;
		break;
	case AM_ZP:
		val = peek(arg % 256);
		break;
	case AM_ZP_X:
		val = peek((arg + regs.x) % 256);
		break;
	case AM_ZP_Y:
		val = peek((arg + regs.y) % 256);
		break;
	case AM_ABS:
		val = peek16(arg);
		break;
	case AM_ABS_X:
		val = peek16(arg + regs.x);
		break;
	case AM_ABS_Y:
		val = peek16(arg + regs.y);
		break;
	case AM_IND_X:
		val = peek(peek((arg + regs.x) % 256) + peek((arg + regs.x + 1) % 256) * 256);
		break;
	case AM_IND_Y:
		val = peek(peek(arg) + peek((arg + 1) % 256) * 256 + regs.y);
		break;
	default:
		fprintf(stderr, "INVALID ADDRESSING MODE\n");
		abort();
	}

	return val;
}

static uint16_t
opcode_mem(enum addressing_mode mode)
{
	uint8_t arg;
	uint16_t val;

	if (mode != AM_ABS && mode != AM_ABS_X && mode != AM_ABS_Y)
		arg = peek(regs.pc++);
	else
		arg = peek16(regs.pc), regs.pc += 2;

	switch (mode) {
	case AM_ZP:
		val = arg % 256;
		break;
	case AM_ZP_X:
		val = (arg + regs.x) % 256;
		break;
	case AM_ZP_Y:
		val = (arg + regs.y) % 256;
		break;
	case AM_ABS:
		val = arg;
		break;
	case AM_ABS_X:
		val = arg + regs.x;
		break;
	case AM_ABS_Y:
		val = arg + regs.y;
		break;
	case AM_IND_X:
		val = peek((arg + regs.x) % 256) + peek((arg + regs.x + 1) % 256) * 256;
		break;
	case AM_IND_Y:
		val = peek(arg) + peek((arg + 1) % 256) * 256 + regs.y;
		break;
	default:
		fprintf(stderr, "INVALID ADDRESSING MODE\n");
		abort();
	}

	return val;
}

static void
adc(uint8_t arg)
{
	uint16_t sum;	// 16-bit sum makes it easier to determine carry flag

	sum = regs.a + arg + regs.status.carry;
	regs.a = sum & 0xFF;

	regs.status.carry = sum > 0xFF;
	/* overflow flag formula: https://stackoverflow.com/a/29224684 */
	regs.status.overflow = (~(regs.a ^ arg) & (regs.a ^ sum) & 0x80) != 0;
	STATUS_UPDATE_ZERO(regs.a);
	STATUS_UPDATE_NEGATIVE(regs.a);
}

static void
and(uint8_t arg)
{
	regs.a &= arg;

	STATUS_UPDATE_ZERO(regs.a);
	STATUS_UPDATE_NEGATIVE(regs.a);
}

static void
asl_acc(void)
{
	uint16_t tmp;

	tmp = regs.a << 1;
	regs.a = tmp & 0xFF;

	regs.status.carry = tmp > 0xFF;
	STATUS_UPDATE_ZERO(regs.a);
	STATUS_UPDATE_NEGATIVE(regs.a);
}

static void
asl(uint16_t mem)
{
	uint16_t tmp;

	tmp = peek(mem) << 1;
	memwrite(mem, tmp & 0xFF);

	regs.status.carry = tmp > 0xFF;
	STATUS_UPDATE_ZERO(tmp);
	STATUS_UPDATE_NEGATIVE(tmp);
}

static void
bcc(uint8_t arg)
{
	if (regs.status.carry == 0)
		regs.pc += arg;
}

static void
bcs(uint8_t arg)
{
	if (regs.status.carry == 1)
		regs.pc += arg;
}

static void
beq(uint8_t arg)
{
	if (regs.status.zero == 1)
		regs.pc += arg;
}

static void
bit(uint8_t arg)
{
	regs.status.zero = (regs.a & arg) == 0;
	regs.status.overflow = (arg & (1 << 6)) != 0;
	STATUS_UPDATE_NEGATIVE(arg);
}

static void
bmi(uint8_t arg)
{
	if (regs.status.negative == 1)
		regs.pc += arg;
}

static void
bne(uint8_t arg)
{
	if (regs.status.zero == 0)
		regs.pc += arg;
}

static void
bpl(uint8_t arg)
{
	if (regs.status.negative == 0)
		regs.pc += arg;
}

static void
brk(void)
{
	/* TODO: push regs.pc and regs.status to stack and load IRQ vector */
	regs.status.brk = 1;
	exit(0);
}

static void
bvc(uint8_t arg)
{
	if (regs.status.overflow == 0)
		regs.pc += arg;
}

static void
bvs(uint8_t arg)
{
	if (regs.status.overflow == 1)
		regs.pc += arg;
}

static void
clc(void)
{
	regs.status.carry = 0;
}

static void
cld(void)
{
	regs.status.decimal_mode = 0;
}

static void
cli(void)
{
	regs.status.interrupt_disable = 0;
}

static void
clv(void)
{
	regs.status.overflow = 0;
}

static void
cmp(uint8_t arg)
{
	uint8_t tmp;

	tmp = regs.a - arg;

	regs.status.carry = regs.a >= arg;
	regs.status.zero = regs.a == arg;
	STATUS_UPDATE_NEGATIVE(tmp);
}

static void
cpx(uint8_t arg)
{
	uint8_t tmp;

	tmp = regs.x - arg;

	regs.status.carry = regs.x >= arg;
	regs.status.zero = regs.x == arg;
	STATUS_UPDATE_NEGATIVE(tmp);
}

static void
cpy(uint8_t arg)
{
	uint8_t tmp;

	tmp = regs.y - arg;

	regs.status.carry = regs.y >= arg;
	regs.status.zero = regs.y == arg;
	STATUS_UPDATE_NEGATIVE(tmp);
}

static void
dec(uint16_t mem)
{
	memwrite(mem, peek(mem) - 1);

	STATUS_UPDATE_ZERO(peek(mem));
	STATUS_UPDATE_NEGATIVE(peek(mem));
}

static void
dex(void)
{
	regs.x--;

	STATUS_UPDATE_ZERO(regs.x);
	STATUS_UPDATE_NEGATIVE(regs.x);
}

static void
dey(void)
{
	regs.y--;

	STATUS_UPDATE_ZERO(regs.y);
	STATUS_UPDATE_NEGATIVE(regs.y);
}

static void
eor(uint8_t arg)
{
	regs.a ^= arg;

	STATUS_UPDATE_ZERO(regs.a);
	STATUS_UPDATE_NEGATIVE(regs.a);
}

static void
inc(uint16_t mem)
{
	memwrite(mem, peek(mem) + 1);

	STATUS_UPDATE_ZERO(peek(mem));
	STATUS_UPDATE_NEGATIVE(peek(mem));
}

static void
inx(void)
{
	regs.x++;

	STATUS_UPDATE_ZERO(regs.x);
	STATUS_UPDATE_NEGATIVE(regs.x);
}

static void
iny(void)
{
	regs.y++;

	STATUS_UPDATE_ZERO(regs.y);
	STATUS_UPDATE_NEGATIVE(regs.y);
}

static void
jmp(uint16_t arg)
{
	regs.pc = arg;
}

static void
jsr(uint16_t arg)
{
	uint16_t tmp = regs.pc + 2;

	/*
	 * first push high-byte of return address then low-byte
	 * https://www.masswerk.at/6502/6502_instruction_set.html
	 */
	PUSH((tmp & 0xFF00) >> 8);
	PUSH(tmp & 0xFF);
	regs.pc = arg;
}

static void
lda(uint8_t arg)
{
	regs.a = arg;

	STATUS_UPDATE_ZERO(regs.a);
	STATUS_UPDATE_NEGATIVE(regs.a);
}

static void
ldx(uint8_t arg)
{
	regs.x = arg;

	STATUS_UPDATE_ZERO(regs.x);
	STATUS_UPDATE_NEGATIVE(regs.x);
}

static void
ldy(uint8_t arg)
{
	regs.y = arg;

	STATUS_UPDATE_ZERO(regs.y);
	STATUS_UPDATE_NEGATIVE(regs.y);
}

static void
lsr_acc(void)
{
	regs.status.carry = regs.a & 1;	// bit 0 in carry
	regs.a >>= 1;
	regs.a &= ~(1 << 7);		// bit 7 cleared

	STATUS_UPDATE_ZERO(regs.a);
	STATUS_UPDATE_NEGATIVE(regs.a);
}

static void
lsr(uint16_t mem)
{
	uint8_t tmp;

	tmp = peek(mem);

	regs.status.carry = tmp & 1;	// bit 0 in carry
	tmp >>= 1;
	tmp &= ~(1 << 7);		// bit 7 cleared

	memwrite(mem, tmp);

	STATUS_UPDATE_ZERO(tmp);
	STATUS_UPDATE_NEGATIVE(tmp);
}

static void
nop(void)
{
	return;
}

static void
ora(uint8_t arg)
{
	regs.a |= arg;

	STATUS_UPDATE_ZERO(regs.a);
	STATUS_UPDATE_NEGATIVE(regs.a);
}

static void
pha(void)
{
	PUSH(regs.a);
}

static void
php(void)
{
	uint8_t status;

	status = (regs.status.carry << 7) | (regs.status.zero << 6)
	    | (regs.status.interrupt_disable << 5) | (regs.status.decimal_mode << 4)
	    | (regs.status.brk << 3) | (regs.status.unused << 2)
	    | (regs.status.overflow << 1) | regs.status.negative;

	PUSH(status);
}

static void
pla(void)
{
	regs.a = PULL();

	STATUS_UPDATE_ZERO(regs.a);
	STATUS_UPDATE_NEGATIVE(regs.a);
}

static void
plp(void)
{
	uint8_t status;

	status = PULL();

	regs.status.carry = (status & (1 << 7)) != 0;
	regs.status.zero = (status & (1 << 6)) != 0;
	regs.status.interrupt_disable = (status & (1 << 5)) != 0;
	regs.status.decimal_mode = (status & (1 << 4)) != 0;
	regs.status.brk = (status & (1 << 3)) != 0;
	regs.status.unused = (status & (1 << 2)) != 0;
	regs.status.overflow = (status & (1 << 1)) != 0;
	regs.status.negative = (status & 1) != 0;
}

static void
rol_acc(void)
{
	uint8_t carry;
	carry = (regs.a & (1 << 7)) != 0;

	regs.a <<= 1;
	regs.a |= regs.status.carry;

	regs.status.carry = carry;
	STATUS_UPDATE_ZERO(regs.a);
	STATUS_UPDATE_NEGATIVE(regs.a);
}

static void
rol(uint16_t mem)
{
	uint8_t carry, tmp;
	carry = (peek(mem) & (1 << 7)) != 0;

	tmp = (peek(mem) << 1) | regs.status.carry;
	memwrite(mem, tmp);

	regs.status.carry = carry;
	STATUS_UPDATE_ZERO(tmp);
	STATUS_UPDATE_NEGATIVE(tmp);
}

static void
ror_acc(void)
{
	uint8_t carry;
	carry = regs.a & 1;

	regs.a >>= 1;
	regs.a |= regs.status.carry << 7;

	regs.status.carry = carry;
	STATUS_UPDATE_ZERO(regs.a);
	STATUS_UPDATE_NEGATIVE(regs.a);
}

static void
ror(uint16_t mem)
{
	uint8_t carry, tmp;
	carry = peek(mem) & 1;

	tmp = (peek(mem) >> 1) | (regs.status.carry << 7);
	memwrite(mem, tmp);

	regs.status.carry = carry;
	STATUS_UPDATE_ZERO(tmp);
	STATUS_UPDATE_NEGATIVE(tmp);
}

static void
rti(void)
{
	plp();
	regs.pc = PULL();
}

static void
rts(void)
{
	regs.pc = PULL() + 1;
}

static void
sbc(uint8_t arg)
{
	/* SBC is described online as ADC with argument as two's complement */
	adc(~arg + 1);
}

static void
sec(void)
{
	regs.status.carry = 1;
}

static void
sed(void)
{
	regs.status.decimal_mode = 1;
}

static void
sei(void)
{
	regs.status.interrupt_disable = 1;
}

static void
sta(uint16_t mem)
{
	memwrite(mem, regs.a);
}

static void
stx(uint16_t mem)
{
	memwrite(mem, regs.x);
}

static void
sty(uint16_t mem)
{
	memwrite(mem, regs.y);
}

static void
tax(void)
{
	regs.x = regs.a;

	STATUS_UPDATE_ZERO(regs.x);
	STATUS_UPDATE_NEGATIVE(regs.x);
}

static void
tay(void)
{
	regs.y = regs.a;

	STATUS_UPDATE_ZERO(regs.y);
	STATUS_UPDATE_NEGATIVE(regs.y);
}

static void
tsx(void)
{
	regs.x = PULL();

	STATUS_UPDATE_ZERO(regs.x);
	STATUS_UPDATE_NEGATIVE(regs.x);
}

static void
txa(void)
{
	regs.a = regs.x;

	STATUS_UPDATE_ZERO(regs.a);
	STATUS_UPDATE_NEGATIVE(regs.a);
}

static void
txs(void)
{
	PUSH(regs.x);
}

static void
tya(void)
{
	regs.a = regs.y;

	STATUS_UPDATE_ZERO(regs.a);
	STATUS_UPDATE_NEGATIVE(regs.a);
}

static void
interpret(void)
{
	uint8_t opcode;

	for (;;) {
		opcode = peek(regs.pc++);

		printf("opcode: $%02X\n", opcode);

		switch (opcode) {
		case 0x69:
			adc(opcode_arg(AM_IMM));
			cycles += 2;
			break;
		case 0x65:
			adc(opcode_arg(AM_ZP));
			cycles += 3;
			break;
		case 0x75:
			adc(opcode_arg(AM_ZP_X));
			cycles += 4;
			break;
		case 0x6d:
			adc(opcode_arg(AM_ABS));
			cycles += 4;
			break;
		case 0x7d:
			adc(opcode_arg(AM_ABS_X));
			cycles += 4;
			break;
		case 0x79:
			adc(opcode_arg(AM_ABS_Y));
			cycles += 4;
			break;
		case 0x72:
			adc(opcode_arg(AM_IND));
			cycles += 5;
			break;
		case 0x61:
			adc(opcode_arg(AM_IND_X));
			cycles += 6;
			break;
		case 0x71:
			adc(opcode_arg(AM_IND_Y));
			cycles += 5;
			break;
		case 0x29:
			and(opcode_arg(AM_IMM));
			cycles += 2;
			break;
		case 0x25:
			and(opcode_arg(AM_ZP));
			cycles += 3;
			break;
		case 0x35:
			and(opcode_arg(AM_ZP_X));
			cycles += 4;
			break;
		case 0x2d:
			and(opcode_arg(AM_ABS));
			cycles += 4;
			break;
		case 0x3d:
			and(opcode_arg(AM_ABS_X));
			cycles += 4;
			break;
		case 0x39:
			and(opcode_arg(AM_ABS_Y));
			cycles += 4;
			break;
		case 0x32:
			and(opcode_arg(AM_IND));
			cycles += 5;
			break;
		case 0x21:
			and(opcode_arg(AM_IND_X));
			cycles += 6;
			break;
		case 0x31:
			and(opcode_arg(AM_IND_Y));
			cycles += 5;
			break;
		case 0x0a:
			asl_acc();
			cycles += 2;
			break;
		case 0x06:
			asl(opcode_arg(AM_ZP));
			cycles += 5;
			break;
		case 0x16:
			asl(opcode_arg(AM_ZP_X));
			cycles += 6;
			break;
		case 0x0e:
			asl(opcode_arg(AM_ABS));
			cycles += 6;
			break;
		case 0x1e:
			asl(opcode_arg(AM_ABS_X));
			cycles += 6;
			break;
		case 0x90:
			bcc(opcode_arg(AM_IMM));
			cycles += 2;
			break;
		case 0xb0:
			bcs(opcode_arg(AM_IMM));
			cycles += 2;
			break;
		case 0xf0:
			beq(opcode_arg(AM_IMM));
			cycles += 2;
			break;
		case 0x89:
			bit(opcode_arg(AM_IMM));
			cycles += 2;
			break;
		case 0x24:
			bit(opcode_arg(AM_ZP));
			cycles += 3;
			break;
		case 0x34:
			bit(opcode_arg(AM_ZP_X));
			cycles += 4;
			break;
		case 0x2c:
			bit(opcode_arg(AM_ABS));
			cycles += 4;
			break;
		case 0x3c:
			bit(opcode_arg(AM_ABS_X));
			cycles += 4;
			break;
		case 0x30:
			bmi(opcode_arg(AM_IMM));
			cycles += 2;
			break;
		case 0xd0:
			bne(opcode_arg(AM_IMM));
			cycles += 2;
			break;
		case 0x10:
			bpl(opcode_arg(AM_IMM));
			cycles += 2;
			break;
		case 0x00:
			brk();
			cycles += 7;
			break;
		case 0x50:
			bvc(opcode_arg(AM_IMM));
			cycles += 2;
			break;
		case 0x70:
			bvs(opcode_arg(AM_IMM));
			cycles += 2;
			break;
		case 0x18:
			clc();
			cycles += 2;
			break;
		case 0xd8:
			cld();
			cycles += 2;
			break;
		case 0x58:
			cli();
			cycles += 2;
			break;
		case 0xb8:
			clv();
			cycles += 2;
			break;
		case 0xc9:
			cmp(opcode_arg(AM_IMM));
			cycles += 2;
			break;
		case 0xc5:
			cmp(opcode_arg(AM_ZP));
			cycles += 3;
			break;
		case 0xd5:
			cmp(opcode_arg(AM_ZP_X));
			cycles += 4;
			break;
		case 0xcd:
			cmp(opcode_arg(AM_ABS));
			cycles += 4;
			break;
		case 0xdd:
			cmp(opcode_arg(AM_ABS_X));
			cycles += 4;
			break;
		case 0xd9:
			cmp(opcode_arg(AM_ABS_Y));
			cycles += 4;
			break;
		case 0xd2:
			cmp(opcode_arg(AM_IND));
			cycles += 5;
			break;
		case 0xc1:
			cmp(opcode_arg(AM_IND_X));
			cycles += 6;
			break;
		case 0xd1:
			cmp(opcode_arg(AM_IND_Y));
			cycles += 5;
			break;
		case 0xe0:
			cpx(opcode_arg(AM_IMM));
			cycles += 2;
			break;
		case 0xe4:
			cpx(opcode_arg(AM_ZP));
			cycles += 3;
			break;
		case 0xec:
			cpx(opcode_arg(AM_ABS));
			cycles += 4;
			break;
		case 0xc0:
			cpy(opcode_arg(AM_IMM));
			cycles += 2;
			break;
		case 0xc4:
			cpy(opcode_arg(AM_ZP));
			cycles += 3;
			break;
		case 0xcc:
			cpy(opcode_arg(AM_ABS));
			cycles += 4;
			break;
		case 0xc6:
			dec(opcode_mem(AM_ZP));
			cycles += 5;
			break;
		case 0xd6:
			dec(opcode_mem(AM_ZP_X));
			cycles += 6;
			break;
		case 0xce:
			dec(opcode_mem(AM_ABS));
			cycles += 6;
			break;
		case 0xde:
			dec(opcode_mem(AM_ABS_X));
			cycles += 7;
			break;
		case 0xca:
			dex();
			cycles += 2;
			break;
		case 0x88:
			dey();
			cycles += 2;
			break;
		case 0x49:
			eor(opcode_arg(AM_IMM));
			cycles += 2;
			break;
		case 0x45:
			eor(opcode_arg(AM_ZP));
			cycles += 3;
			break;
		case 0x55:
			eor(opcode_arg(AM_ZP_X));
			cycles += 4;
			break;
		case 0x4d:
			eor(opcode_arg(AM_ABS));
			cycles += 4;
			break;
		case 0x5d:
			eor(opcode_arg(AM_ABS_X));
			cycles += 4;
			break;
		case 0x59:
			eor(opcode_arg(AM_ABS_Y));
			cycles += 4;
			break;
		case 0x52:
			eor(opcode_arg(AM_IND));
			cycles += 5;
			break;
		case 0x41:
			eor(opcode_arg(AM_IND_X));
			cycles += 6;
			break;
		case 0x51:
			eor(opcode_arg(AM_IND_Y));
			cycles += 5;
			break;
		case 0xe6:
			inc(opcode_arg(AM_ZP));
			cycles += 5;
			break;
		case 0xf6:
			inc(opcode_arg(AM_ZP_X));
			cycles += 6;
			break;
		case 0xee:
			inc(opcode_arg(AM_ABS));
			cycles += 6;
			break;
		case 0xfe:
			inc(opcode_arg(AM_ABS_X));
			cycles += 7;
			break;
		case 0xe8:
			inx();
			cycles += 2;
			break;
		case 0xc8:
			iny();
			cycles += 2;
			break;
		case 0x4c:
			jmp(opcode_arg(AM_ABS));
			cycles += 3;
			break;
		case 0x6c:
			jmp(opcode_arg(AM_IND));
			cycles += 6;
			break;
		case 0x7c:
			jmp(opcode_arg(AM_ABS_X));
			cycles += 6;
			break;
		case 0x20:
			jsr(opcode_arg(AM_ABS));
			cycles += 6;
			break;
		case 0xa9:
			lda(opcode_arg(AM_IMM));
			cycles += 2;
			break;
		case 0xa5:
			lda(opcode_arg(AM_ZP));
			cycles += 3;
			break;
		case 0xb5:
			lda(opcode_arg(AM_ZP_X));
			cycles += 4;
			break;
		case 0xad:
			lda(opcode_arg(AM_ABS));
			cycles += 4;
			break;
		case 0xbd:
			lda(opcode_arg(AM_ABS_X));
			cycles += 4;
			break;
		case 0xb9:
			lda(opcode_arg(AM_ABS_Y));
			cycles += 4;
			break;
		case 0xb2:
			lda(opcode_arg(AM_IND));
			cycles += 5;
			break;
		case 0xa1:
			lda(opcode_arg(AM_IND_X));
			cycles += 6;
			break;
		case 0xb1:
			lda(opcode_arg(AM_IND_Y));
			cycles += 5;
			break;
		case 0xa2:
			ldx(opcode_arg(AM_IMM));
			cycles += 2;
			break;
		case 0xa6:
			ldx(opcode_arg(AM_ZP));
			cycles += 3;
			break;
		case 0xb6:
			ldx(opcode_arg(AM_ZP_Y));
			cycles += 4;
			break;
		case 0xae:
			ldx(opcode_arg(AM_ABS));
			cycles += 4;
			break;
		case 0xbe:
			ldx(opcode_arg(AM_ABS_Y));
			cycles += 4;
			break;
		case 0xa0:
			ldy(opcode_arg(AM_IMM));
			cycles += 2;
			break;
		case 0xa4:
			ldy(opcode_arg(AM_ZP));
			cycles += 3;
			break;
		case 0xb4:
			ldy(opcode_arg(AM_ZP_X));
			cycles += 4;
			break;
		case 0xac:
			ldy(opcode_arg(AM_ABS));
			cycles += 4;
			break;
		case 0xbc:
			ldy(opcode_arg(AM_ABS_X));
			cycles += 4;
			break;
		case 0x4a:
			lsr_acc();
			cycles += 2;
			break;
		case 0x46:
			lsr(opcode_arg(AM_ZP));
			cycles += 5;
			break;
		case 0x56:
			lsr(opcode_arg(AM_ZP_X));
			cycles += 6;
			break;
		case 0x4e:
			lsr(opcode_arg(AM_ABS));
			cycles += 6;
			break;
		case 0x5e:
			lsr(opcode_arg(AM_ABS_X));
			cycles += 6;
			break;
		case 0xea:
			nop();
			cycles += 2;
			break;
		case 0x09:
			ora(opcode_arg(AM_IMM));
			cycles += 2;
			break;
		case 0x05:
			ora(opcode_arg(AM_ZP));
			cycles += 3;
			break;
		case 0x15:
			ora(opcode_arg(AM_ZP_X));
			cycles += 4;
			break;
		case 0x0d:
			ora(opcode_arg(AM_ABS));
			cycles += 4;
			break;
		case 0x1d:
			ora(opcode_arg(AM_ABS_X));
			cycles += 4;
			break;
		case 0x19:
			ora(opcode_arg(AM_ABS_Y));
			cycles += 4;
			break;
		case 0x12:
			ora(opcode_arg(AM_IND));
			cycles += 5;
			break;
		case 0x01:
			ora(opcode_arg(AM_IND_X));
			cycles += 6;
			break;
		case 0x11:
			ora(opcode_arg(AM_IND_Y));
			cycles += 5;
			break;
		case 0x48:
			pha();
			cycles += 3;
			break;
		case 0x08:
			php();
			cycles += 3;
			break;
		case 0x68:
			pla();
			cycles += 4;
			break;
		case 0x28:
			plp();
			cycles += 4;
			break;
		case 0x2a:
			rol_acc();
			cycles += 2;
			break;
		case 0x26:
			rol(opcode_arg(AM_ZP));
			cycles += 5;
			break;
		case 0x36:
			rol(opcode_arg(AM_ZP_X));
			cycles += 6;
			break;
		case 0x2e:
			rol(opcode_arg(AM_ABS));
			cycles += 6;
			break;
		case 0x3e:
			rol(opcode_arg(AM_ABS_X));
			cycles += 6;
			break;
		case 0x6a:
			ror_acc();
			cycles += 2;
			break;
		case 0x66:
			ror(opcode_arg(AM_ZP));
			cycles += 5;
			break;
		case 0x76:
			ror(opcode_arg(AM_ZP_X));
			cycles += 6;
			break;
		case 0x6e:
			ror(opcode_arg(AM_ABS));
			cycles += 6;
			break;
		case 0x7e:
			ror(opcode_arg(AM_ABS_X));
			cycles += 6;
			break;
		case 0x40:
			rti();
			cycles += 6;
			break;
		case 0x60:
			rts();
			cycles += 6;
			break;
		case 0xe9:
			sbc(opcode_arg(AM_IMM));
			cycles += 2;
			break;
		case 0xe5:
			sbc(opcode_arg(AM_ZP));
			cycles += 3;
			break;
		case 0xf5:
			sbc(opcode_arg(AM_ZP_X));
			cycles += 4;
			break;
		case 0xed:
			sbc(opcode_arg(AM_ABS));
			cycles += 4;
			break;
		case 0xfd:
			sbc(opcode_arg(AM_ABS_X));
			cycles += 4;
			break;
		case 0xf9:
			sbc(opcode_arg(AM_ABS_Y));
			cycles += 4;
			break;
		case 0xf2:
			sbc(opcode_arg(AM_IND));
			cycles += 5;
			break;
		case 0xe1:
			sbc(opcode_arg(AM_IND_X));
			cycles += 6;
			break;
		case 0xf1:
			sbc(opcode_arg(AM_IND_Y));
			cycles += 5;
			break;
		case 0x38:
			sec();
			cycles += 2;
			break;
		case 0xf8:
			sed();
			cycles += 2;
			break;
		case 0x78:
			sei();
			cycles += 2;
			break;
		case 0x85:
			sta(opcode_mem(AM_ZP));
			cycles += 4;
			break;
		case 0x95:
			sta(opcode_mem(AM_ZP_X));
			cycles += 5;
			break;
		case 0x8d:
			sta(opcode_mem(AM_ABS));
			cycles += 5;
			break;
		case 0x9d:
			sta(opcode_mem(AM_ABS_X));
			cycles += 6;
			break;
		case 0x99:
			sta(opcode_mem(AM_ABS_Y));
			cycles += 6;
			break;
		case 0x92:
			sta(opcode_mem(AM_IND));
			cycles += 6;
			break;
		case 0x81:
			sta(opcode_mem(AM_IND_X));
			cycles += 7;
			break;
		case 0x91:
			sta(opcode_mem(AM_IND_Y));
			cycles += 7;
			break;
		case 0x86:
			stx(opcode_mem(AM_ZP));
			cycles += 4;
			break;
		case 0x96:
			stx(opcode_mem(AM_ZP_Y));
			cycles += 5;
			break;
		case 0x8e:
			stx(opcode_mem(AM_ABS));
			cycles += 5;
			break;
		case 0x84:
			sty(opcode_mem(AM_ZP));
			cycles += 4;
			break;
		case 0x94:
			sty(opcode_mem(AM_ZP_X));
			cycles += 5;
			break;
		case 0x8c:
			sty(opcode_mem(AM_ABS));
			cycles += 5;
			break;
		case 0xaa:
			tax();
			cycles += 2;
			break;
		case 0xa8:
			tay();
			cycles += 2;
			break;
		case 0xba:
			tsx();
			cycles += 2;
			break;
		case 0x8a:
			txa();
			cycles += 2;
			break;
		case 0x9a:
			txs();
			cycles += 2;
			break;
		case 0x98:
			tya();
			cycles += 2;
			break;
		default:
			printf("opcode $%02X not implemented\n", opcode);
			break;
		}

		printf("status: %i%i%i%i%i%i%i%i\n", regs.status.carry,
		    regs.status.zero, regs.status.interrupt_disable,
		    regs.status.decimal_mode, regs.status.brk,
		    regs.status.unused, regs.status.overflow,
		    regs.status.negative);
		printf("A: $%02X, X: $%02X, Y: $%02X, SP: $%02X, PC: $%02X\n",
		    regs.a, regs.x, regs.y, regs.sp, regs.pc);
	}
}

/* https://www.nesdev.org/wiki/CPU_power_up_state */
void
cpu_init(void)
{
	regs.a = regs.x = regs.y = 0;
	regs.pc = 0xFFFC;
	regs.sp = 0xFD;

	memset(&regs.status, 0, sizeof(regs.status));
	regs.status.unused = 1;

	cycles += 7;
}

int
main(void)
{
	cpu_init();

	if (sizeof(program) > (0x10000 - 0x8000)) {
		fprintf(stderr, "program is too big for memory\n");
		return 1;
	}

	memcpy(memory + 0x8000, program, sizeof(program));
	regs.pc = 0x8000;

	printf("Initial State:\n");
	printf("status: %i%i%i%i%i%i%i%i\n", regs.status.carry,
	    regs.status.zero, regs.status.interrupt_disable,
	    regs.status.decimal_mode, regs.status.brk,
	    regs.status.unused, regs.status.overflow,
	    regs.status.negative);
	printf("A: $%02X, X: $%02X, Y: $%02X, SP: $%02X, PC: $%02X\n",
	    regs.a, regs.x, regs.y, regs.sp, regs.pc);

	putchar('\n');

	interpret();

	return 0;
}