AMD 250 manual 121

25112 Rev. 3.06 September 2005 Software Optimization Guide for AMD64 Processors

before starting a copy, especially for large blocks. To write data directly to main memory, bypassing the cache, use the MOVNTI instruction instead of MOV for the four store instructions.

;rsi = source

;rdi = destination

;ecx = byte count

mov	eax, ecx
shr	eax, 5
jz	done_32
align 16		; align the loop to a 16-byte fetch boundary
copy_32_bytes:
mov	r8, [rsi]	; read 8 bytes
mov	r9, [rsi+8]	; it's a bit faster to pair two reads
add	rsi, 32	; update source pointer
mov	[rdi], r8	;	store 8 bytes
mov	[rdi+8], r9	; again, pair 2 stores for slight perf gain
add	rdi, 32	;	update destination pointer
mov	r8, [rsi-16]	; loop is unrolled 4 reads, 4 writes
mov	r9, [rsi-8]	; 4-way unroll hides latency of adds and dec
dec	eax	; decrement data counter (32 bytes)
mov	[rdi-16], r8	;	store more bytes
mov	[rdi-8], r9	; store last 8 bytes
jnz	copy_32_bytes
done_32:

(copy any remaining bytes)

Chapter 5 Cache and Memory Optimizations 121