str1 db 'abcdefghijklmnopqrstuvwxyzabc',0
str2 db 64 dup(0)

	lea esi,str1	 ; load effective address of string 
	lea edi,str2	 ; load effective address of buffer
	mov ecx,[sizeof str1-1] ; determine length of string less terminating '0'
	mov ebx,ecx	 ; load length for manipulation
	cmp ecx,4	 ; ensure string is large enough for dword manipulations
	jb l1		 ; if not, work with one byte at a time
	and ebx,3	 ; test dword alignment by checking two least significant bytes
	jz l2		 ; if zero it's 4 byte aligned
	cmp ebx,3	 ; test if both lsb's set
	jnz @f		 ; next test
	sub ecx,1	 ; both lsb's are set, work byte by byte until dword aligned
	mov al,[esi+ecx] ; load last character
	sub ebx,1	 ; decrement for next test
	mov [edi],al     ; save byte to beginning of buffer
	add edi,1	 ; increment buffer position for next byte

@@:	cmp ebx,2	 ; test only bit 1
	jnz @f		 ; next test
	sub ecx,1	 ; decrement string position
	mov al,[esi+ecx] ; load byte
	sub ebx,1	 ; prepare for next byte
	mov [edi],al	 ; save byte to buffer
	add edi,1	 ; increment buffer position

@@:	cmp ebx,1	 ; test only bit 0
	jnz l1		 ; next
	sub ecx,1	 ; update string position
	mov al,[esi+ecx] ; load byte
	mov [edi],al	 ; store to buffer
	add edi,1	 ; update buffer position
	jmp l2		 ; move on to dword manipulation
	
l1:     sub ecx,1        ; string is less than 4 characters
	mov al,[esi+ecx] ; load byte
	mov [edi],al     ; store byte
	add edi,1        ; update buffer
	or ecx,ecx	 ; last byte?
	jz l3		 ; done
	jmp l1	         ; continue
	
l2:     sub ecx,4        ; back up 4 bytes
	jnz @f	         ; if > 0 go to dword loading loop
	mov eax,[esi+ecx]; else load the only dword
	bswap eax        ; reverse dword 
	mov [edi],eax    ; store dword
	jmp l3	         ; done
	   
@@:     mov eax,[esi+ecx]; load dword 
	bswap eax	 ; swap bytes
	mov [edi],eax    ; store dword
	add edi,4	 ; update buffer for next dword
	sub ecx,4	 ; back up to next dword
	jnz @b	         ; loop
	mov eax,[esi]    ; load last dword
	bswap eax	 ; swap
	mov [edi],eax    ; store and done. str2 contains str1 reversed 
l3: