[llvm] [X86][AVX] Prefer per-element vector shifts for known splats #39424 (PR #87913)

[Simon Pilgrim via llvm-commits]

RKSimon wrote:

Only AVX512 targets can broadcast from GPRs directly - AVX2 has to move it on to the vector registers first (usually via a ISD:SCALAR_TO_VECTOR node - which lowers in this case to vmovd)

https://github.com/llvm/llvm-project/pull/87913