[llvm] [X86][LV] Add partial-reduction dot-product support (PR #205373)

[Florian Hahn via llvm-commits]

https://github.com/fhahn requested changes to this pull request.

@hhy3 the backend changes are completely independent of LV support and should be a separate PR.

> LoopVectorize currently caps candidate VFs using the widest element type seen in the loop. Dot-product reductions widen narrow inputs, such as i8, i16, or bf16, into i32/f32 recurrence values before the reduction update. Treating the widened recurrence type as the loop's effective widest type can keep the selected VF too small even when the target can lower the recurrence as a partial reduction.


Right, this should also be done as separate PR as follow-up after initial support. 

https://github.com/llvm/llvm-project/pull/205373