[LV] missed optimization with pragma assume_safety

[vfdff]

• I defined 2 function Map_Center_Of_Mass_CPU and foo_noalias, we can see the only different is that the function foo_noalias has extra restrict for the argument int * crd. According Control-auto-vectorization-with-pragmas, the #pragma clang loop vectorize(assume_safety) can assume there is no data dependencies for the following loop, so I expected they have similar output assemble , and now only the function foo_noalias can vectorize succcessfully.

The preceding pragma indicates to the compiler that the following loop contains no data dependencies between loop iterations that would prevent vectorization. The compiler might be able to use this information to vectorize a loop, where it would not typically be possible

• test: https://gcc.godbolt.org/z/sjThWPs9f

void Map_Center_Of_Mass_CPU(int residue_i, int *start,  int *end, int * crd)
{
    #pragma clang loop vectorize(assume_safety)
    for (int atom_i = start[residue_i]; atom_i < end[residue_i]; atom_i++)
    {
        crd[atom_i] = crd[atom_i] + 8;
    }
}


void foo_noalias (int residue_i, int *start,  int *end, int * __restrict__ crd)
{
    #pragma clang loop vectorize(assume_safety)
    for (int atom_i = start[residue_i]; atom_i < end[residue_i]; atom_i++)
    {
        crd[atom_i] = crd[atom_i] + 8;
    }
}

[Fangtangtang]

Rpass remarks that the loop is not vectorized because it could not determine the number of loop iterations. I think this is due to its inability to decide whether there exists an alias between end and crd, and thus it cannot parallelize the iterations.

[vfdff]

Thanks for your comment.

As the #pragma clang loop vectorize(assume_safety) can assume there is no data dependencies for the following loop, does it mean there is no dependencies between end and crd, .ie they are not alias ?

[Fangtangtang]

I found that it can be vectorized if I load the value before the loop and use this value in loop condition.

    int end_value = end[residue_i];
    #pragma clang loop vectorize(assume_safety)
    for (int atom_i = start[residue_i]; atom_i < end_value; atom_i++)
    {
        crd[atom_i] = crd[atom_i] + 8;
    }

Also, if I update end_value with end[residue_i] in the loop body, it can not be vectorized as well.

    #pragma clang loop vectorize(assume_safety)
    for (int atom_i = start[residue_i]; atom_i < end_value; atom_i++)
    {
        crd[atom_i] = crd[atom_i] + 8;
        end_value = end[residue_i];
    }

• test: https://godbolt.org/z/4sdYoPT43

My personal guess is that #pragma clang loop vectorize(assume_safety) indicates to the compiler that the loop contains no data dependencies between iterations that would prevent vectorization. However, there could be a potential dependency between end[residue_i] and crd[atom_i] within a single iteration.

You might want to refer to the source code for more detailed information. There are several return statements before IsAnnotatedParallel skips the memory dependence checks, as seen in this section of the LLVM project. It's likely that one of those checks failed in this case, preventing vectorization.
I'm not very familiar with this part, so apologies for not exploring it further.