Справочник Пользователя для AMD 250

25112

Rev. 3.06

September 2005

❖

For a very fast general purpose memory copy routine, call the libc memcpy() function included 

with the Microsoft or gcc tools.  This function features optimizations for all block sizes and 
alignments.

This optimization applies to:

•

32-bit software

•

64-bit software

The memcpy() routines included with recent compilers from Microsoft and gcc feature optimizations 
for all block sizes and alignments for AMD Athlon 64 and AMD Opteron processors.

Use inline assembly code to copy a small data structure in cache. Use an unrolled series of MOV 
instructions. Alternate loads and stores in sequences such as load/store/load/store routines, or use 
load/load/store/store sequences for even better performance. Align the destination (and source) if 
possible.

The following 64-bit example copies 18 bytes of data:

; rsi = source

; rdi = destination

    mov     r8, [rsi]      ; 8 bytes of source

    mov     r9, [rsi+8]    ; next 8 bytes of source

    mov     [rdi], r8      ; write 8 bytes

    mov     [rdi+8], r9    ; write next 8

    mov     r8w, [rsi+16]  ; read two bytes "r8 word"

    mov     [rdi+16], r8w  ; write the last 2 bytes

The following example illustrates how to copy blocks of 32 bytes and larger, in cache. This code 
performs best when the source and destination addresses are 8-byte aligned. Align the destination