Can counting byte matches between two strings be optimized using SIMD?

Question

Profiling suggests that this function here is a real bottle neck for my application:

static inline int countEqualChars(const char* string1, const char* strin         


        

Answer 1

Auto-vectorization in current gcc is a matter of helping the compiler to understand that's easy to vectorize the code. In your case: it will understand the vectorization request if you remove the conditional and rewrite the code in a more imperative way:

    static inline int count(const char* string1, const char* string2, int size) {
            int r = 0;
            bool b;

            for (int j = 0; j < size; ++j) {
                    b = (string1[j] == string2[j]);
                    r += b;
            }

            return r;
    }

In this case:

movdqa  16(%rsp), %xmm1
movl    $.LC2, %esi
pxor    %xmm2, %xmm2
movzbl  416(%rsp), %edx
movdqa  .LC1(%rip), %xmm3
pcmpeqb 224(%rsp), %xmm1
cmpb    %dl, 208(%rsp)
movzbl  417(%rsp), %eax
movl    $1, %edi
pand    %xmm3, %xmm1
movdqa  %xmm1, %xmm5
sete    %dl
movdqa  %xmm1, %xmm4
movzbl  %dl, %edx
punpcklbw   %xmm2, %xmm5
punpckhbw   %xmm2, %xmm4
pxor    %xmm1, %xmm1
movdqa  %xmm5, %xmm6
movdqa  %xmm5, %xmm0
movdqa  %xmm4, %xmm5
punpcklwd   %xmm1, %xmm6

(etc.)