Mutual reductions not vectorized in SLP

[davemgreen]

Given this code:
https://godbolt.org/z/G55n584nb

uint64_t both(uint32_t* x) {
    uint32_t s = 0;
    uint32_t m = 0;
    for(int i = 0; i < 32; i++) {
        s += x[i];
        m += x[i] * x[i];
    }
    return s + ((uint64_t)m << 32);
}

If we remove the s += line then it vectorizes the MLA. If we remove the m += line then it vectorizes the Add. But both together it does not.

[labrinea]

SLP seems to bail because of the extraction cost for the external users of each reduction. My understanding is that when calculating the extraction cost for one of the reductions, there's no vectorization tree representing the other reduction yet, and so SLP thinks it's not profitable to vectorize.

Patch: https://reviews.llvm.org/D133441

[alexey-bataev]

It is currently vectorized, not sure it produces the best result

[davemgreen]

I think we get the costs wrong and end up accidentally get it right at the moment. #138811 breaks things again.