[clang] [llvm] [AggressiveInstCombine] Shrink loads used in shufflevector rebroadcasts. (PR #128938)

[Nikita Popov via llvm-commits]

https://github.com/nikic commented:

This needs to be a cost-model driven transform in VectorCombine. Shrinking vector loads is not always profitable.

https://github.com/llvm/llvm-project/pull/128938