vectorization in icc

Hi,


I've been experimenting with the Intel C/C++ compiler for Linux, and in
particular, with the automatic vectorization.


I have the following piece of code (all the arrays are of type char):


for( j = 0; j < 16; j++ ) {


                d[ j ] = d[ j ] + d[ j ];
                d[ j ] = d[ j ] | B[ j ];
                dm[ j ] = d[ j ] & mm[ j ];
}




which compiles to the following, if I disable the vectorization:


..B3.16: # Preds ..B3.16 ..B3.15
                movb 4656(%esp,%ecx), %al #178.23
                movb 4672(%esp,%ecx), %dl #181.23
                addb %al, %al #178.23
                orb (%edi,%ecx), %al #179.23
                movb %al, 4656(%esp,%ecx) #179.4
                andb %dl, %al #181.23
                movb %al, 4688(%esp,%ecx) #181.4
                addl $1, %ecx #176.23
                cmpl $16, %ecx #176.3
                jl ..B3.16 # Prob 93% #176.3




With vectorization enabled, I get the following, i.e. the loop is
eliminated by using sse2 instructions:


                paddb %xmm1, %xmm1 #177.14
                por 80(%esp,%ecx,8), %xmm1 #178.14
                movdqa %xmm1, %xmm3 #180.14
                lea 1(%edi), %eax #183.44
                addl $1, %esi #169.21
                pand %xmm0, %xmm3 #180.14
                movdqa %xmm3, 4720(%esp) #180.4


Both work just fine, but the vectorized code is significantly slower!
I certainly expected the vectorized code to be much faster.


What's going on? Are the sse2 instructions really so slow compared to
the standard integer instructions? If so, what's the point of the
vectorization anyways?


Thanks.

Return to the comp.compilers page.
Search the comp.compilers archives again.