Check whether __m128i is zero?

Question

I found this question:

Which I used to create the below example:

int main(){

    __m128i intrinreg;
    intrinreg.m128i_i64[0] = 0;
    intrinreg.m128i_i64[1] = 6;

    __m128i zero = _mm_setzero_si128();

    if (0 == _mm_movemask_epi8(_mm_cmpeq_epi32(intrinreg, zero)))
    {
        std::cout << "true" << std::endl;
    }
    else{
        std::cout << "false" << std::endl;
    }
}

but whether I set the two 64-bit int components to 0 or a non-zero, I always get "false" print out?

If the value is zero, then obviously the comparison against zero can't result in zero. Right now it prints "true" if neither element is zero. — harold, Jun 07 '15 at 18:07
If the above returns whether all 128 bits are zero, then alternating the '6' to '0' should change the output, yeah? — user997112, Jun 07 '15 at 18:29
The accepted answer given for the other question doesn't actually answer the question asked in the other question. The expression evaluates to true only if all the 32-bit integer components are not equal to zero. If you want the expression to evaluate to true if all 64-bit components are zero you need to use the expression given by Leonid Tsybert in his comment: `_mm_movemask_epi8(_mm_cmpeq_epi32(x,zero)) == 0xFFFF` You should also use `_mm_cmpeq_epi64` instead of `_epi32` but in this case you get the same result. — Ross Ridge, Jun 07 '15 at 18:31
[Check XMM register for all zeroes](http://stackoverflow.com/q/10175711/995714) — phuclv, Jun 08 '15 at 07:57

score 2 · Answer 1 · answered Jun 08 '15 at 07:41

Since you have tagged AVX I assume that you have SSE4.1 in which case the instruction you want is ptest which you can get from _mm_testz_si128 or _mm_test_all_zeros().

bool test1, test2;
__m128i r1 = _mm_setzero_si128();
__m128i r2 = _mm_set1_epi64x(42);
test1 = _mm_testz_si128(r1,r1);
test2 = _mm_testz_si128(r2,r2);
printf("%d %d\n", test1, test2); //prints 1 0

If you don't have SSE4.1 then use _mm_movemask_epi8(_mm_cmpeq_epi32(x,_mm_setzero_si128())) == 0xFFFF. This requires pcmpeqd, pmovmskb, and test.

However, ptest sets the zero flag (ZF) so it only needs one instruction. See checking-if-two-sse-registers-are-not-both-zero-without-destroying-them for more details.

Check whether __m128i is zero?

1 Answers1