[llvm] [X86][BF16] Improve vectorization of BF16 (PR #88486)

[Phoebe Wang via llvm-commits]

phoebewang wrote:

> LGTM - although I think eventually it'd be nice to get bf16->f32 "zext+shift" out of X86 and handle it in generic legalization

Thanks @RKSimon, the main point here is to combine "FP_ROUND+FP_EXTEND". It's a bit hard if we lower to "zext+shift" and then combine "FP_ROUND+zext+shift".

https://github.com/llvm/llvm-project/pull/88486