x86-64 Assembly: Tutorial - A Quick Guide to the Changes in 64-bit Assembly - Page 3

Copy Memory Using 64-bit Registers

This x64 assembly code will copy memory from one place to another using 8-byte registers then falls back to using 1-byte registers for the remainder. For performance reasons, no division operations are performed. Notice that the stack is never touched here.

_TEXT	SEGMENT
memcopyx64 PROC

	mov rax, rcx            ; move the dest pointer in rcx to rax
	mov rcx, r8             ; move length to rcx
	shr rcx, 3				; divide by 8 - holds quotient now
	shl rcx, 3				; divide by 8 - holds quotient now
	sub r8, rcx             ; find the remainder and store in r8
	shr rcx, 3				; divide by 8 - holds quotient now

	cmp rcx, 0              ; make sure there is at least 8 bytes to copy
	je compare_remainder

copy8bytes:
	mov r9, QWORD PTR [rdx] ; copy 8 bytes from src to r9
	mov QWORD PTR [rax], r9 ; copy 8 bytes from r9 to dest
	add rax, 8
	add rdx, 8
	loopnz copy8bytes       ; loop while rcx > 0 - this will automatically decrement rcx

compare_remainder:
	cmp r8, 0               ; if there is no remainder then finished
	je exit

	mov rcx, r8             ; move the remainder to rcx

copybytes:
	mov r9b, BYTE PTR [rdx] ; copy 1 byte from src to r9b
	mov BYTE PTR [rax], r9b ; copy 1 byte from r9b to dest
	inc rax
	inc rdx
	loopnz copybytes        ; loop while rcx > 0 - this will automatically decrement rcx

exit:
	ret

memcopyx64 ENDP
_TEXT	ENDS
END

The driver source code:

extern "C" char * memcopyx64(char *dest, const char *src, unsigned long long length);

int main()
{
	char dest[256];
	const char* src = "abcdefghijklmnopqrstuvwxyz!ABCDEFGHIJKLMNOPQRSTUVWXYZ!";
	unsigned int len = 0;
	while (src[len])
		len++;
	len++; // add the sentinel to the length
	char* end = dest;
	for (int i = 0; i < 3; i++)
		end = memcopyx64(end, src, len) - 1; // subtract the sentinel
	return 0;
}

Copy Memory Using 128-bit Registers

This x64 assembly code will copy memory from one place to another using 16-byte registers then falls back to using 2-byte registers for the remainder with the exception of the last odd byte. For performance reasons, no division operations are performed just as above, and the stack is never touched here. This code is a port of the x86 version.

_TEXT	SEGMENT
memcopy128 PROC

	mov rax, rcx                   ; move the dest pointer in rcx to rax

	mov r10, r8                    ; move length to r10
	shr r10, 4				       ; divide by 16 - holds quotient now
	shl r10, 4                     ; multiply by 16
	sub r8, r10                    ; find the remainder and store in r8
	shr r10, 4				       ; divide by 16 - holds quotient now

	cmp r10, 0                     ; make sure there are at least 16 bytes to copy
	je compare_remainder

	mov rcx, r10                   ; move the quotient to rcx

copy16bytes:
	movdqu xmm0, XMMWORD PTR [rdx] ; copy 16 bytes from src to xmm0
	movdqu XMMWORD PTR [rax], xmm0 ; copy 16 bytes from xmm0 to dest
	add rax, 16
	add rdx, 16
	loopnz copy16bytes             ; loop while rcx > 0 - this will automatically decrement rcx

compare_remainder:
	cmp r8, 0                      ; if there is no remainder then finished
	je exit

	mov rcx, r8                    ; move the remainder to rcx
	shr rcx, 1				       ; divide by 2
	cmp rcx, 0                     ; make sure there are at least 2 bytes to copy
	je check_odd_byte

copy2bytes:
	mov r9w, WORD PTR [rdx]        ; copy 2 bytes from src to r9w
	mov WORD PTR [rax], r9w        ; copy 2 byte from r9w to dest
	add rax, 2
	add rdx, 2
	loopnz copy2bytes              ; loop while rcx > 0 - this will automatically decrement rcx

check_odd_byte:
	test r8, 1
	jz exit

	mov r9b, BYTE PTR [rdx]        ; copy 1 byte from src to r9b
	mov BYTE PTR [rax], r9b        ; copy 1 byte from r9b to dest
	inc rax

exit:
	ret

memcopy128 ENDP
_TEXT	ENDS
END

The driver source code:

extern "C" char * memcopy128(char *dest, const char *src, unsigned long long length);

int main()
{
	char dest[256];
	const char* src = "abcdefghijklmnopqrstuvwxyz!ABCDEFGHIJKLMNOPQRSTUVWXYZ!";
	unsigned int len = 0;
	while (src[len])
		len++;
	len++; // add the sentinel to the length
	char* end = dest;
	for (int i = 0; i < 3; i++)
		end = memcopy128(end, src, len) - 1; // subtract the sentinel
	return 0;
}

---