[llvm] [VectorCombine] Add subvector reduction support to foldShuffleChainsToReduce (PR #199872)

[Adam Scott via llvm-commits]

as4230 wrote:

@RKSimon Ready for review and sorry for the churn. I rebased onto trunk so it handles FADD/FMUL. I also reworked the matcher to be top-down instead of bottom-up and dropped the poison-refinement case to keep it simpler. Instead of propagating bitmasks forward through the chain it propagates counts backward. It should be easier to walk through now.


https://github.com/llvm/llvm-project/pull/199872