AMD Typewriter x86 사용자 설명서

22007E/0—November 1999

   /* block fill (destination QWORD aligned) */

   __asm {

      mov       edx, [dst_ptr]
      mov       ecx, [blk_size]
      shr       ecx, 6
      movq      mm0, [fill_data]
     
      align 16

   $fill_nc:
      movntq    [edx], mm0
      movntq    [edx+8], mm0
      movntq    [edx+16], mm0
      movntq    [edx+24], mm0
      movntq    [edx+32], mm0
      movntq    [edx+40], mm0
      movntq    [edx+48], mm0
      movntq    [edx+56], mm0
      add       edx, 64
      dec       ecx
      jnz       $fill_nc
      femms
      sfence
   }

To clear all the bits in an MMX register to zero, use:

PXOR  MMreg, MMreg

Note that PXOR MMreg, MMreg is dependent on previous
writes to MMreg. Therefore, using PXOR in the manner
described can lengthen dependency chains, which in return
may lead to reduced performance. An alternative in such cases
is to use:

zero DD 0

MOVD  MMreg, DWORD PTR [zero]

i.e., to load a zero from a statically initialized and properly
aligned memory location. However, loading the data from
memory runs the risk of cache misses. Cases where MOVD is
superior to PXOR are therefore rare and PXOR should be used
in general.