clang loop vectorizer doesn't work when reduction variable is pointer type

[llvmbot]

Bugzilla Link	27625
Version	3.8
OS	Linux
Reporter	LLVM Bugzilla Contributor
CC	@hfinkel

Extended Description

#define SIZE (1 << 15)

#define XMM_ALIGNMENT_BYTES 16

float *vec_a __attribute__((aligned (XMM_ALIGNMENT_BYTES)));
float *vec_b __attribute__((aligned (XMM_ALIGNMENT_BYTES)));
float *vec_c __attribute__((aligned (XMM_ALIGNMENT_BYTES)));

void matvec_autovec()
{
    int i, j;

    for (i = 0; i < SIZE; i++) {
        for (j = 0; j < SIZE; j++) {
            vec_c[i] += vec_a[j] * vec_b[j];
        }
    }
}

Command used: clang-3.8 -c -S -Ofast -Rpass-analysis=loop-vectorize test.c

test.c:17:22: remark: loop not vectorized: value that could not be identified as reduction is used outside the loop [-Rpass-analysis=loop-vectorize]
            vec_c[i] += vec_a[j] * vec_b[j];```

I was expecting that vec_c[i] to be treated as inner loop invariant and the vectorization would work. The auto-vectorization works fine if I use a tmp variable for accumulation. Could someone explain why the vectorizer bails out in this case?

[llvmbot]

float vec_a[SIZE] __attribute__((aligned (XMM_ALIGNMENT_BYTES)));
float vec_b[SIZE] __attribute__((aligned (XMM_ALIGNMENT_BYTES)));
float vec_c[SIZE] __attribute__((aligned (XMM_ALIGNMENT_BYTES)));

Further testing shows that the vectorizer works if I use array as shown above. Changed the title accordingly as well.

[fhahn]

This is now gets vectorized https://clang.godbolt.org/z/q5M3vax47

Likely due to 4e5e042