[llvm] [SLP]Initial support for (masked)loads + compress and (masked)interleaved (PR #132099)

[Alexey Bataev via llvm-commits]

alexey-bataev wrote:

> FWIW, it looks like rewriting this as a loop would make the code slightly simpler:
> 
> ```
> @@ -20797,10 +20797,10 @@
>            CandidateVFs.push_back(Limit);
>          if (VF > MaxTotalNum || VF >= StoresLimit)
>            break;
> -        for_each(RangeSizes, [&](std::pair<unsigned, unsigned> &P) {
> +        for(auto& P : RangeSizes) {
>            if (P.first != 0)
>              P.first = std::max(P.second, P.first);
> -        });
> +        }
>          // Last attempt to vectorize max number of elements, if all previous
>          // attempts were unsuccessful because of the cost issues.
>          CandidateVFs.push_back(VF);
> ```
> 
> or
> 
> ```
> @@ -20797,10 +20797,10 @@
>            CandidateVFs.push_back(Limit);
>          if (VF > MaxTotalNum || VF >= StoresLimit)
>            break;
> -        for_each(RangeSizes, [&](std::pair<unsigned, unsigned> &P) {
> +        for(std::pair<unsigned, unsigned> &P : RangeSizes) {
>            if (P.first != 0)
>              P.first = std::max(P.second, P.first);
> -        });
> +        }
>          // Last attempt to vectorize max number of elements, if all previous
>          // attempts were unsuccessful because of the cost issues.
>          CandidateVFs.push_back(VF);
> ```
> 
> And it also makes the crash disappear, though it may be due to the issue being fragile.

Hmm, semantically these 2 pieces are equal. Maybe a bug in the loop vectorizer, that causes memory pollution on data rewriting?  Did you compile clang itself with another clang? If so, what is the version?

https://github.com/llvm/llvm-project/pull/132099