[llvm] [LV] Convert uniform-address unmasked scatters to scalar store. (PR #166114)

[Luke Lau via llvm-commits]

lukel97 wrote:

> Narrow this patch to only convert unmasked scatter to scalar store for not blocking. Rebase on top of #166514 to show the test changes.
> 
> > This is a cool VPlan transformation, but I think in most of the changed tests here LICM will already pull out these uniform stores before it hits the loop vectorizer. Were you seeing any workloads or benchmarks where this prevented scatters?
> 
> I tried to apply this patch in our downstream compiler on GeekBench6 and it showed the dynamic instruction count difference. But I haven't collected the performance data yet.

I tried the latest version of this patch out on SPEC CPU 2017 and llvm-test-suite with -march=rva23u64 -O3 but there was no difference in the output. Was there a specific benchmark on geekbench that you saw the improvement on that we can try to recreate this with upstream?

https://github.com/llvm/llvm-project/pull/166114