AVR Freaks

If I build:

#include <avr/io.h>

uint8_t a,b;

int main(void)
{
	a = a - b;
	while(1) {
	}
}

The .LSS file shows:

	a = a - b;
    80d4:	80 91 01 20 	lds	r24, 0x2001	; 0x802001 <a>
    80d8:	90 91 00 20 	lds	r25, 0x2000	; 0x802000 <__DATA_REGION_ORIGIN__>
    80dc:	89 1b       	sub	r24, r25
    80de:	80 93 01 20 	sts	0x2001, r24	; 0x802001 <a>

The bytes of the SUB opcode are 0x89, 0x1B but this is little endian so the 16 bit opcode is actually 0x1B89. The opcode manual says:

so if you write 0x1B89 as 0001 1011 1000 1001 then the 0001 10 at the start are fixed and tell you this is "SUB" but then the r and d bits contains the encoding of the specific registers used (r24, r25). rrrrr is given by the 3rd bit in the second nibble then all the bits of the 4th nibble so it is 1 1001. 0b11001 is 25 so that is the encoding for R25. The d bits are the last bit of the 2nd nibble and the four bits of the third nibble so is 1 1000. 0b11000 is 24. So that is the encoding of R24.

The complete opcode is 0x1B89.

All AVR opcodes are some multiple of 16 bits. More complex instructions like JMP might take two lots of 16 bits for example. When the AVR makes an opcode fetch it reads TWO consecutive bytes from memory. This is actually why Atmel have a rather nasty habit of always referring to flash memory in terms or word size/addressing not byte addressing. So in a 32K (bytes) 328P they would tell you the flash is 16K words. That's because the most opcodes you could get into it (assuming you picked to use only single word opcodes) would be 16384 operations.