[llvm] [LLVM][AArch64] Prefer zip(0,A) over shll(A) when promoting bfloat vectors. (PR #187540)

[Paul Walker via llvm-commits]

paulwalker-arm wrote:

> Hi - Do you have any real benchmarks to check that the results of llvm-mca are in the right ballpark?

Nothing specific but I'll see if I can write something.

> These patterns are a bit strange - they were added without code review and should have been written very differently. They seem to effectively be doing type legalizating during selection. Do you think it is worth redoing them?

I agree but wasn't sure where they came from.  I'll reimplement their lowering as a separate PR and circle back to this one later.



https://github.com/llvm/llvm-project/pull/187540