[llvm] [LV]Reject narrow VFs for narrow FP reductions fed by gappy interleaves (PR #222422)

[Alexey Bataev via llvm-commits]

alexey-bataev wrote:

> Is this only an issue for FP reductions?
Maybe not only for reductions, but in our case it is reductions-related, would be good to check for other cases too.

> If the issue is wasting of bandwidth through the loads, should the interleave group cost be adjusted?
It is not about costing, it tries to address the memory bandwidth. For our case, even llvm-mca says that the vector code should be about 1,7 x faster, but the actual perf is about 4.4 x slower because of the unmodelled bandwidth

https://github.com/llvm/llvm-project/pull/222422