How to populate a 64 bit register with duplicate byte values

Question

I\'m doing some x64 assembly with Visual C++ 2010 and masm (\'fast call\' calling convention).

So let\'s say I have a function in C++:

extern \"C\" v         


        

Answer 1

Because you called your procedure 'fillArray', I assumed you like to fill a whole memory block with a byte value. So I did a comparision on different approaches. It is 32 bit masm code, but the results should be similar in 64 bit mode. Each approach is tested with both aligned and unaligned buffers. Here are the results:

Simple REP STOSB - aligned....: 192
Simple REP STOSB - not aligned: 192
Simple REP STOSD - aligned....: 191
Simple REP STOSD - not aligned: 222
Simple while loop - aligned....: 267
Simple while loop - not aligned: 261
Simple while loop with different addressing - aligned....: 271
Simple while loop with different addressing - not aligned: 262
Loop with 16-byte SSE write - aligned....: 192
Loop with 16-byte SSE write - not aligned: 205
Loop with 16-byte SSE write non-temporal hint - aligned....: 126 (EDIT)

The most naive variant using the following code seems to perform best in both scenarios and has the smallest code size as well:

cld
mov al, 44h   ; byte value
mov edi, lpDst
mov ecx, 256000*4  ; buf size
rep stosb

EDIT: It's not the fastest for aligned data. Added MOVNTDQ version which performs best, see below.

For the sake of completeness, here are excerpts from the other routines - the value is assumed to be expanded into EAX before:

Rep Stosd:

mov edi, lpDst
mov ecx, 256000
rep stosd

Simple While:

mov edi, lpDst
mov ecx, 256000
.while ecx>0
    mov [edi],eax
    add edi,4
    dec ecx
.endw

Different simple while:

mov edi, lpDst
xor ecx, ecx
.while ecx<256000 
    mov [edi+ecx*4],eax
    inc ecx
.endw

SSE(both):

movd xmm0,eax
punpckldq xmm0,xmm0    ; xxxxxxxxGGGGHHHH -> xxxxxxxxHHHHHHHH
punpcklqdq xmm0,xmm0   ; xxxxxxxxHHHHHHHH -> HHHHHHHHHHHHHHHH
mov ecx, 256000/4   ; 16 byte
mov edi, lpDst
.while ecx>0 
    movdqa xmmword ptr [edi],xmm0    ; movdqu for unaligned
    add edi,16
    dec ecx
.endw

SSE(NT,aligned,EDIT):

movd xmm0,eax
punpckldq xmm0,xmm0    ; xxxxxxxxGGGGHHHH -> xxxxxxxxHHHHHHHH
punpcklqdq xmm0,xmm0   ; xxxxxxxxHHHHHHHH -> HHHHHHHHHHHHHHHH
mov ecx, 256000/4   ; 16 byte
mov edi, lpDst
.while ecx>0 
    movntdq xmmword ptr [edi],xmm0
    add edi,16
    dec ecx
.endw

I uploaded the whole code here http://pastie.org/9831404 --- the MASM package from hutch is required for assembling.

---

If SSSE3 is available, you can use pshufb to broadcast a byte to all positions of a register instead of a chain of punpck instructions.

movd    xmm0, edx
xorps   xmm1,xmm1      ; xmm1 = 0
pshufb  xmm0, xmm1     ; xmm0 = _mm_set1_epi8(dl)