[llvm] [CodeGen] Expansion of scalable vector reductions (PR #129214)

[Matt Arsenault via llvm-commits]

arsenm wrote:

]> Expanding into a single loop would require a lot of changes in `PreISelIntrinsicLowering` I assume? How about running a stripped-down and simplified version of something like the LoopFuse pass on demand inside of `PreISelIntrinsicLowering` (but only if scalable vector intrinsics have been expanded)? That might later be easier to bring to non-intrinsic scalable operations needing expansion than trying to create a single loop from the start?

Sounds beyond the scope of PreISelIntrinsicLowering. If you wanted a fancier pass, it should probably be separate. It might still be useful to have the support PreISelIntrinsicLowering as the simple fallback (e.g. the fused lowering would be an optional optimization, with the simple lowering fallback) 

https://github.com/llvm/llvm-project/pull/129214