[llvm] [AArch64] SVE Shuffleopt: merge reduction reverse into tbl (PR #206047)

[Paul Walker via llvm-commits]

https://github.com/paulwalker-arm commented:

Something to consider is whether the transform specifically needs to see a tbl. I'm wondering if the core idea of isolating shuffles to a single binop operand has any downsides?

I mention this because if it doesn't then you could implement this top-down from the partial reduction as a generic inst combine prior to the shuffle pass, which would then generate a suitable shuffle mask when it sees the extra vector_reverse.

https://github.com/llvm/llvm-project/pull/206047