Gigabit integer add optimization

[qWord]

my suggestion would be:

Code Select Expand

CHUNK_INFO struct
    pChunkA     PVOID   ? ; in , destination
    pChunkB     PVOID   ? ; in , source
    n           DWORD   ? ; in , number of QWORDs
    carry       DWORD   ? ; in/out  , 0 or 1
CHUNK_INFO ends


; thread procedure
BlockAdd proc pChunkInfo:ptr CHUNK_INFO
   
    mov rbx,pChunkInfo
    mov ecx,[rbx].CHUNK_INFO.n
    mov rdi,[rbx].CHUNK_INFO.pChunkA
    mov rsi,[rbx].CHUNK_INFO.pChunkB
    shr ecx,3
    mov edx,[rbx].CHUNK_INFO.carry
    align 16
@@:
    mov r8 ,[rsi+0*8]
    mov r9 ,[rsi+1*8]
    mov r10,[rsi+2*8]
    mov r11,[rsi+3*8]
    mov r12,[rsi+4*8]
    mov r13,[rsi+5*8]
    mov r14,[rsi+6*8]
    mov r15,[rsi+7*8]
    lea rsi,[rsi+8*8]
    add r8,rdx
    mov rdx,0
    adc [rdi+0*8],r8
    adc [rdi+1*8],r9
    adc [rdi+2*8],r10
    adc [rdi+3*8],r11
    adc [rdi+4*8],r12
    adc [rdi+5*8],r13
    adc [rdi+6*8],r14
    adc [rdi+7*8],r15
    adc rdx,0
    lea rdi,[rdi+8*8]
   
    ; this loop could be unrolled several times ...
   
    dec ecx
    jnz @B
   
    mov [rbx].CHUNK_INFO.carry,edx
   
    mov rax,1
    ret
   
BlockAdd endp

The structure is filled before calling CreateThread() with the needed addresses. Also a possible carry is returned through it.

EDIT: code changed

[jj2007]

PADDQ doesn't do carries.

Valid argument. Now if your data are predictable, you could branch into SSE code if no carry expected (test for bit 63?). Usually the speed gain with SSE2 is considerable...