[llvm] [VectorCombine] Expand `vector_insert` into shufflevector for earlier cost optimizations (#145512) (PR #146479)

[Alexey Bataev via llvm-commits]

alexey-bataev wrote:

> Given vector_insert always gets expanded in instcombine a couple of passes later and its causing problems with missed folds in vectorcombine - could SLP can just emit shuffles?

Sure, we have a wrapper, that emits insert_vector, if possible, otherwise falls back to the regular shuffle. If shuffles are preferred, I can remove emission of the insert_vector

https://github.com/llvm/llvm-project/pull/146479