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

[via llvm-commits]

laurenmchin wrote:

> Why is this necessary? Is there some pass that emits fixed vector insert instead of shufflevector?

Yes, because SLPVectorizer runs before VectorCombine in the default O2/O3 pipelines. 

check-llvm tests like `alternate-cast.ll` expect `vector_insert` to be folded to `shufflevector` before SLP runs, which only happens if the fold is in InstCombine, otherwise SLP won’t catch it. You can see this in `PassBuilderPipelines.cpp`, in `PassBuilder::addVectorPasses`: InstCombinePass -> SLPVectorizerPass -> VectorCombinePass.

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