.data
.align 32
;#define B0 (0x55555555)
B0:
.dw 0x55555555
;#define B1 (0x33333333)
B1:
.dw 0x33333333
;#define B2 (0x0F0F0F0F)
B2:
.dw 0x0F0F0F0F
;#define B3 (0x001F001F)
B3:
.dw 0x001F001F
;#define B4 (0x0000003F)
B4:
.dw 0x0000003F

.text

;int countbits(int x) {
countbits:
	;x == r1
	
	;load address of B0 into r3
	ldil r3,low(B0);
	ldih r3,high(B0);

	;x = (x & B0) + ((x >>  1) & B0);
	;r0 = (x >> 1)
	lrs r0,r1,1;
	
	;load B0 into r2
	ldx r2, r3;

	;(x >> 1) & B0	
	and r0, r0, r2;

	;x & B0
	and r1, r1, r2;
	
	;x = (x & B0) + ((x >>  1) & B0);
	add r0, r0, r1;

	;load B1 into r2, use displacment for address of B1 (or add?)
	ldx r2, (r3)(B1-B0);

	;x = (x & B1) + ((x >>  2) & B1);
	lrs r1, r0, 2;
	and r1, r1, r2;

	and r0, r0, r2;

	add r0, r0, r1;

	;x = (x + (x >>  4)) & B2;
	ldx r2, (r3)(B2-B0);

	lrs r1, r0, 4;
	and r1, r1, r2;
	and r0, r0, r2;

	add r0, r0, r1;

	;x = (x + (x >>  8)) /*& B3 */
	lrs r1, r0, 8;
	add r0, r0, r1;

	ldx r2, (r3)(B4-B0);
	;x = (x + (x >> 16)) & B4;
	lrs r1, r0, 16;
	add r0, r0, r1;
	and r0, r0, r2;

	;return x;
	ret;
;}

int computeparity(int x) 
	x ^= x >> 16;
	x ^= x >> 8;
	x ^= x >> 4;
	x &= 0xf;
	return (0x6996 >> x) & 1;
}

void main(void) {
	countbits(B0);
	computeparity(B0);
	computeparity(B0+1);
}