/* Macros for setting/clearing program status register bits */

#define F_setN(v)	if (v) P |= 0x80; else P &= ~0x80
#define F_setV(v)	if (v) P |= 0x40; else P &= ~0x40
#define F_setM(v)	if (v) P |= 0x20; else P &= ~0x20
#define F_setX(v)	if (v) P |= 0x10; else P &= ~0x10
#define F_setD(v)	if (v) P |= 0x08; else P &= ~0x08
#define F_setI(v)	if (v) P |= 0x04; else P &= ~0x04
#define F_setZ(v)	if (v) P |= 0x02; else P &= ~0x02
#define F_setC(v)	if (v) P |= 0x01; else P &= ~0x01

/* Macros for testing program status register bits */

#define F_getN	((P & 0x80)? 1:0)
#define F_getV	((P & 0x40)? 1:0)
#define F_getM	((P & 0x20)? 1:0)
#define F_getX	((P & 0x10)? 1:0)
#define F_getD	((P & 0x08)? 1:0)
#define F_getI	((P & 0x04)? 1:0)
#define F_getZ	((P & 0x02)? 1:0)
#define F_getC	((P & 0x01)? 1:0)
#define F_getE	(E? 1:0)

/*------- Routines that operate on an 8/16-bit value  -------*/

#define C_SETF8(v)	F_setN(v & 0x80);	\
			F_setZ(!v)

#define C_SETF16(v)	F_setN(v & 0x8000);	\
			F_setZ(!v)

#define C_LDA8(v)	A.B.L = v;	\
			C_SETF8(v)

#define C_LDA16(v)	A.W = v;	\
			C_SETF16(v)

#define C_LDX8(v)	X.B.L = v;	\
			C_SETF8(v)

#define C_LDX16(v)	X.W = v;	\
			C_SETF16(v)

#define C_LDY8(v)	Y.B.L = v;	\
			C_SETF8(v)

#define C_LDY16(v)	Y.W = v;	\
			C_SETF16(v)

#define C_INC8(v)	v++;		\
			C_SETF8(v)

#define C_INC16(v)	v++;		\
			C_SETF16(v)

#define C_DEC8(v)	v--;		\
			C_SETF8(v)

#define C_DEC16(v)	v--;		\
			C_SETF16(v)

#define C_ASL8(v)	F_setC(v & 0x80);	\
			v = v << 1;		\
			C_SETF8(v)

#define C_ASL16(v)	F_setC(v & 0x8000);	\
			v = v << 1;		\
			C_SETF16(v)

#define C_LSR8(v)	F_setC(v & 0x01);	\
			v = v >> 1;		\
			C_SETF8(v)

#define C_LSR16(v)	F_setC(v & 0x0001);	\
			v = v >> 1;		\
			C_SETF16(v)

#define C_ROL8(v)	wtmp.B.L = P & 0x01;		\
			F_setC(v & 0x80);		\
			v = (v << 1) | wtmp.B.L;	\
			C_SETF8(v)

#define C_ROL16(v)	wtmp.W = P & 0x01;	\
			F_setC(v & 0x8000);	\
			v = (v << 1) | wtmp.W;	\
			C_SETF16(v)

#define C_ROR8(v)	wtmp.B.L = (P & 0x01) << 7;	\
			F_setC(v & 0x01);		\
			v = (v >> 1) | wtmp.B.L;	\
			C_SETF8(v)

#define C_ROR16(v)	wtmp.W = (P & 0x01) << 15;	\
			F_setC(v & 0x0001);		\
			v = (v >> 1) | wtmp.W; 		\
			C_SETF16(v)

#define C_AND8(v)	A.B.L &= v;	\
			C_SETF8(A.B.L)

#define C_AND16(v)	A.W &= v;	\
			C_SETF16(A.W)

#define C_ORA8(v)	A.B.L |= v;	\
			C_SETF8(A.B.L)

#define C_ORA16(v)	A.W |= v;	\
			C_SETF16(A.W)

#define C_EOR8(v)	A.B.L ^= v;	\
			C_SETF8(A.B.L)

#define C_EOR16(v)	A.W ^= v;	\
			C_SETF16(A.W)

#define C_TSB8(v)	F_setN(v & 0x80);	\
			F_setV(v & 0x40);	\
			v |=  A.B.L;		\
			F_setZ(!v)

#define C_TSB16(v)	F_setN(v & 0x8000);	\
			F_setV(v & 0x4000);	\
			v |= A.W;		\
			F_setZ(!v)

#define	C_TRB8(v)	F_setN(v & 0x80);	\
			F_setV(v & 0x40);	\
			v &= ~A.B.L;		\
			F_setZ(!v)

#define C_TRB16(v)	F_setN(v & 0x8000);	\
			F_setV(v & 0x4000);	\
			v &= ~A.W;		\
			F_setZ(!v)

#define C_BIT8(v)	F_setN(v & 0x80);	\
			F_setV(v & 0x40);	\
			F_setZ(!(v & A.B.L));	\

#define C_BIT16(v)	F_setN(v & 0x8000);	\
			F_setV(v & 0x4000);	\
			F_setZ(!(v & A.W));	\

#define C_ADC8(v)	if (F_getD) {							\
				printf("WARNING -- decimal mode ADC attempted!\n");	\
			} else {							\
				wtmp.W = A.B.L + v + F_getC;				\
				F_setV(~(A.B.L ^ v) & (A.B.L ^ wtmp.B.L) & 0x80);	\
				F_setC(wtmp.B.H);					\
				A.B.L = wtmp.B.L;					\
				C_SETF8(A.B.L);						\
			}

#define C_ADC16(v)	if (F_getD) {							\
				printf("WARNING -- decimal mode ADC attempted!\n");	\
			} else {							\
				atmp.A = A.W + v + F_getC;				\
				F_setV(~(A.W ^ v) & (A.W ^ atmp.A) & 0x8000);		\
				F_setC(atmp.B.B);					\
				A.W = atmp.A;						\
				C_SETF16(A.W);						\
			}

#define C_SBC8(v)	if (F_getD) {							\
				printf("WARNING -- decimal mode SBC attempted!\n");	\
			} else {							\
				wtmp.W = (A.B.L - v) - !F_getC;				\
				F_setV((A.B.L ^ v) & (A.B.L ^ wtmp.B.L) & 0x80);	\
				F_setC(!wtmp.B.H);					\
				A.B.L = wtmp.B.L;					\
				C_SETF8(A.B.L);						\
			}

#define C_SBC16(v)	if (F_getD) {							\
				printf("WARNING -- decimal mode SBC attempted!\n");	\
			} else {							\
				atmp.A = (A.W - v) - !F_getC;				\
				F_setV((A.W ^ v) & (A.W ^ atmp.A) & 0x8000);		\
				F_setC(!atmp.B.B);					\
				A.W = atmp.A;						\
				C_SETF16(A.W);						\
			}

#define C_CMP8(v)	wtmp.W = A.B.L - v;	\
			F_setC(!wtmp.B.H);	\
			C_SETF8(wtmp.B.L)

#define C_CMP16(v)	atmp.A = A.W - v;	\
			F_setC(!atmp.B.B);	\
			atmp.B.B = 0;		\
			C_SETF16(atmp.A)

#define C_CPX8(v)	wtmp.W = X.B.L - v;	\
			F_setC(!wtmp.B.H);	\
			C_SETF8(wtmp.B.L)

#define C_CPX16(v)	atmp.A = X.W - v;	\
			F_setC(!atmp.B.B);	\
			atmp.B.B = 0;		\
			C_SETF16(atmp.A)

#define C_CPY8(v)	wtmp.W = Y.B.L - v;	\
			F_setC(!wtmp.B.H);	\
			C_SETF8(wtmp.B.L)

#define C_CPY16(v)	atmp.A = Y.W - v;	\
			F_setC(!atmp.B.B);	\
			atmp.B.B = 0;		\
			C_SETF16(atmp.A)

/*------- Routines that operate on a memory address -------*/

#define C_LDA(a)	otmp.B.L = M_READ(a.A);			\
			if (F_getM) {				\
				C_LDA8(otmp.B.L);		\
			} else {				\
				otmp.B.H = M_READ(a.A+1);	\
				C_LDA16(otmp.W);		\
			}

#define C_LDX(a)	otmp.B.L = M_READ(a.A);			\
			if (F_getM) {				\
				C_LDX8(otmp.B.L);		\
			} else {				\
				otmp.B.H = M_READ(a.A+1);	\
				C_LDX16(otmp.W);		\
			}

#define C_LDY(a)	otmp.B.L = M_READ(a.A);			\
			if (F_getM) {				\
				C_LDY8(otmp.B.L);		\
			} else {				\
				otmp.B.H = M_READ(a.A+1);	\
				C_LDY16(otmp.W);		\
			}

#define C_STA(a)	M_WRITE(a.A,A.B.L);			\
			if (!F_getM) M_WRITE(a.A+1,A.B.H)

#define C_STX(a)	M_WRITE(a.A,X.B.L);			\
			if (!F_getX) M_WRITE(a.A+1,X.B.H)

#define C_STY(a)	M_WRITE(a.A,Y.B.L);			\
			if (!F_getX) M_WRITE(a.A+1,Y.B.H)

#define C_STZ(a)	M_WRITE(a.A,0);				\
			if (!F_getM) M_WRITE(a.A+1,0)

#define C_INC(a)	otmp.B.L = M_READ(a.A);			\
			if (F_getM) {				\
				C_INC8(otmp.B.L);		\
				M_WRITE(a.A,otmp.B.L);		\
			} else {				\
				otmp.B.H = M_READ(a.A+1);	\
				C_INC16(otmp.W);		\
				M_WRITE(a.A,otmp.B.L);		\
				M_WRITE(a.A+1,otmp.B.H);	\
			}

#define C_DEC(a)	otmp.B.L = M_READ(a.A);			\
			if (F_getM) {				\
				C_DEC8(otmp.B.L);		\
				M_WRITE(a.A,otmp.B.L);		\
			} else {				\
				otmp.B.H = M_READ(a.A+1);	\
				C_DEC16(otmp.W);		\
				M_WRITE(a.A,otmp.B.L);		\
				M_WRITE(a.A+1,otmp.B.H);	\
			}

#define C_ASL(a)	otmp.B.L = M_READ(a.A);			\
			if (F_getM) {				\
				C_ASL8(otmp.B.L);		\
				M_WRITE(a.A,otmp.B.L);		\
			} else {				\
				otmp.B.H = M_READ(a.A+1);	\
				C_ASL16(otmp.W);		\
				M_WRITE(a.A,otmp.B.L);		\
				M_WRITE(a.A+1,otmp.B.H);	\
			}

#define C_LSR(a)	otmp.B.L = M_READ(a.A);			\
			if (F_getM) {				\
				C_LSR8(otmp.B.L);		\
				M_WRITE(a.A,otmp.B.L);		\
			} else {				\
				otmp.B.H = M_READ(a.A+1);	\
				C_LSR16(otmp.W);		\
				M_WRITE(a.A,otmp.B.L);		\
				M_WRITE(a.A+1,otmp.B.H);	\
			}

#define C_ROL(a)	otmp.B.L = M_READ(a.A);			\
			if (F_getM) {				\
				C_ROL8(otmp.B.L);		\
				M_WRITE(a.A,otmp.B.L);		\
			} else {				\
				otmp.B.H = M_READ(a.A+1);	\
				C_ROL16(otmp.W);		\
				M_WRITE(a.A,otmp.B.L);		\
				M_WRITE(a.A+1,otmp.B.H);	\
			}

#define C_ROR(a)	otmp.B.L = M_READ(a.A);			\
			if (F_getM) {				\
				C_ROR8(otmp.B.L);		\
				M_WRITE(a.A,otmp.B.L);		\
			} else {				\
				otmp.B.H = M_READ(a.A+1);	\
				C_ROR16(otmp.W);		\
				M_WRITE(a.A,otmp.B.L);		\
				M_WRITE(a.A+1,otmp.B.H);	\
			}

#define C_AND(a)	otmp.B.L = M_READ(a.A);			\
			if (F_getM) {				\
				C_AND8(otmp.B.L);		\
			} else {				\
				otmp.B.H = M_READ(a.A+1);	\
				C_AND16(otmp.W);		\
			}

#define C_ORA(a)	otmp.B.L = M_READ(a.A);			\
			if (F_getM) {				\
				C_ORA8(otmp.B.L);		\
			} else {				\
				otmp.B.H = M_READ(a.A+1);	\
				C_ORA16(otmp.W);		\
			}

#define C_EOR(a)	otmp.B.L = M_READ(a.A);			\
			if (F_getM) {				\
				C_EOR8(otmp.B.L);		\
			} else {				\
				otmp.B.H = M_READ(a.A+1);	\
				C_EOR16(otmp.W);		\
			}

#define C_TSB(a)	otmp.B.L = M_READ(a.A);			\
			if (F_getM) {				\
				C_TSB8(otmp.B.L);		\
				M_WRITE(a.A,otmp.B.L);		\
			} else {				\
				otmp.B.H = M_READ(a.A+1);	\
				C_TSB16(otmp.W);		\
				M_WRITE(a.A,otmp.B.L);		\
				M_WRITE(a.A+1,otmp.B.H);	\
			}

#define C_TRB(a)	otmp.B.L = M_READ(a.A);			\
			if (F_getM) {				\
				C_TRB8(otmp.B.L);		\
				M_WRITE(a.A,otmp.B.L);		\
			} else {				\
				otmp.B.H = M_READ(a.A+1);	\
				C_TRB16(otmp.W);		\
				M_WRITE(a.A,otmp.B.L);		\
				M_WRITE(a.A+1,otmp.B.H);	\
			}

#define C_BIT(a)	otmp.B.L = M_READ(a.A);			\
			if (F_getM) {				\
				C_BIT8(otmp.B.L);		\
			} else {				\
				otmp.B.H = M_READ(a.A+1);	\
				C_BIT16(otmp.W);		\
			}

#define C_ADC(a)	otmp.B.L = M_READ(a.A);			\
			if (F_getM) {				\
				C_ADC8(otmp.B.L);		\
			} else {				\
				otmp.B.H = M_READ(a.A+1);	\
				C_ADC16(otmp.W);		\
			}

#define C_SBC(a)	otmp.B.L = M_READ(a.A);			\
			if (F_getM) {				\
				C_SBC8(otmp.B.L);		\
			} else {				\
				otmp.B.H = M_READ(a.A+1);	\
				C_SBC16(otmp.W);		\
			}

#define C_CMP(a)	otmp.B.L = M_READ(a.A);			\
			if (F_getM) {				\
				C_CMP8(otmp.B.L);		\
			} else {				\
				otmp.B.H = M_READ(a.A+1);	\
				C_CMP16(otmp.W);		\
			}

#define C_CPX(a)	otmp.B.L = M_READ(a.A);			\
			if (F_getX) {				\
				C_CPX8(otmp.B.L);		\
			} else {				\
				otmp.B.H = M_READ(a.A+1);	\
				C_CPX16(otmp.W);		\
			}

#define C_CPY(a)	otmp.B.L = M_READ(a.A);			\
			if (F_getX) {				\
				C_CPY8(otmp.B.L);		\
			} else {				\
				otmp.B.H = M_READ(a.A+1);	\
				C_CPY16(otmp.W);		\
			}

/* Macros for pushing or pulling bytes on the 65816 stack */

#define S_PUSH(v)	M_WRITE(S.W--,v); if (F_getE) S.B.H = 1;

#define S_PULL(v)	v = M_READ(++S.W); if (F_getE) S.B.H = 1;

/* Macros to retrieve an 8 or 16-bit operand. They take as their parameter	*/
/* a "dualw" union variable, which they set to the operand.			*/

#define O_i8(v)		v.B.L = M_READ(PC.W++); v.B.H = 0

#define O_i16(v)	v.B.L = M_READ(PC.W++); v.B.H = M_READ(PC.W++)

/* Macros to retrieve the operand address. These take as their parameter	*/
/* a "duala" union variable, which they set to the operand address.		*/

#define O_a(a)		a.B.L = M_READ(PC.W++); a.B.H = M_READ(PC.W++); a.B.B = DB

#define O_al(a)		a.B.L = M_READ(PC.W++); a.B.H = M_READ(PC.W++); a.B.B = M_READ(PC.W++)

#define O_d(a)		a.A = D.W + M_READ(PC.W++); a.B.B = 0

#define O_dix(a)	atmp.A = D.W + M_READ(PC.W++); a.B.B = 0;			\
			a.B.L = M_READ(atmp.A); a.B.H = M_READ(atmp.A+1); a.B.B = DB;	\
			a.A += Y.W

#define O_dixl(a)	atmp.A = D.W + M_READ(PC.W++); a.B.B = 0;					\
			a.B.L = M_READ(atmp.A); a.B.H = M_READ(atmp.A+1); a.B.B = M_READ(atmp.A+2);	\
			a.A += Y.W
			
#define O_dxi(a)	atmp.A = D.W + M_READ(PC.W++) + X.W; a.B.B = 0;			\
			a.B.L = M_READ(atmp.A); a.B.H = M_READ(atmp.A+1); a.B.B = DB

#define O_dxx(a)	a.A = (M_READ(PC.W++) + D.W + X.W); a.B.B = 0;	\
			if (F_getE) a.B.H = 0

#define O_dxy(a)	a.A = (M_READ(PC.W++) + D.W + Y.W); a.B.B = 0;	\
			if (F_getE) a.B.H = 0

#define O_axx(a)	a.B.L = M_READ(PC.W++); a.B.H = M_READ(PC.W++); a.B.B = DB; a.A += X.W;	\
			if (F_getE) a.B.B = 0

#define O_axy(a)	a.B.L = M_READ(PC.W++); a.B.H = M_READ(PC.W++); a.B.B = DB; a.A += Y.W;	\
			if (F_getE) a.B.B = 0

#define O_alxx(a)	a.B.L = M_READ(PC.W++); a.B.H = M_READ(PC.W++); a.B.B = M_READ(PC.W++); a.A += X.W

#define O_pcr(a)	atmp.B.L = M_READ(PC.W++);				\
			a.A = PC.W + (offset_s) atmp.B.L; a.B.B = PB;

#define O_pcrl(a)	wtmp.B.L = M_READ(PC.W++); wtmp.B.H = M_READ(PC.W++);	\
			a.A = PC.W + (offset_l) wtmp.W; a.B.B = PB;

#define O_ai(a)		atmp.B.L = M_READ(PC.W++); atmp.B.H = M_READ(PC.W++); atmp.B.B = 0;	\
			a.B.L = M_READ(atmp.A); a.B.H = M_READ(atmp.A+1); a.B.B = 0

#define O_ail(a)	atmp.B.L = M_READ(PC.W++); atmp.B.H = M_READ(PC.W++); atmp.B.B = 0;	\
			a.B.L = M_READ(atmp.A); a.B.H = M_READ(atmp.A+1); a.B.B = M_READ(atmp.A+2)

#define O_di(a)		atmp.A = M_READ(PC.W++) + D.W;					\
			a.B.L = M_READ(atmp.A); a.B.H = M_READ(atmp.A+1); a.B.B = DB

#define O_dil(a)	atmp.A = M_READ(PC.W++) + D.W;					\
			a.B.L = M_READ(atmp.A); a.B.H = M_READ(atmp.A+1); a.B.B = M_READ(atmp.A+2)

#define O_axi(a)	atmp.B.L = M_READ(PC.W++); atmp.B.H = M_READ(PC.W++); atmp.A += X.W; atmp.B.B = 0; \
			a.B.L = M_READ(atmp.A); a.B.H = M_READ(atmp.A+1); a.B.B = 0

#define O_sr(a)		a.A = (M_READ(PC.W++) + S.W) & 0xFFFF;

#define O_srix(a)	atmp.A = (M_READ(PC.W++) + S.W) & 0xFFFF;	\
			a.B.L = M_READ(atmp.A); a.B.H = M_READ(atmp.A+1); a.B.B = DB; a.A += Y.W