x86_mmx_memcpy.asm

   1 bits 32
   2 section .text
   3 ;extern "C" int mmx_memcpy
   4 ;  (unsigned char *dest, unsigned char *src, int len);
   5
   6 global mmx_memcpy
   7
   8 times ($$-$) & 3 db 0
   9
  10 mmx_memcpy:
  11
  12 pushad                  ; save registers
  13 mov edi,[esp+36]        ; get 1st argument
  14 mov esi,[esp+40]        ; ...2nd
  15 mov eax,[esp+44]        ; ...3rd
  16
  17 mov edx, eax
  18 shr eax, byte 3         ; figure out how many 8 byte chunks we have
  19 and edx, byte 7         ; also figure out slack
  20 test eax, eax           ; Do we have any big chunks?
  21 push edx
  22 jz .slack               ; If not, let's just do slack
  23
  24 mov ecx,eax
  25
  26 .mmx_move:
  27   movq mm0,qword[esi]   ; move 8 byte blocks using MMX
  28   movq qword[edi],mm0
  29   add esi, byte 8       ; increment pointers
  30   add edi, byte 8
  31 loopnz .mmx_move        ; continue until CX=0
  32
  33 .slack:
  34 pop ecx
  35 rep movsb               ; move 1 byte slack
  36
  37 emms                    ; Free up for the FPU
  38
  39 popad                   ; clean up
  40 ret
  41
  42 ; --------------------------------------
  43
  44 %ifdef NASM_STACK_NOEXEC
  45 section .note.GNU-stack noalloc noexec nowrite progbits
  46 %endif