[clang] [llvm] [SROA] Canonicalize homogeneous structs into fixed vectors (PR #165159)

[Yaxun Liu via llvm-commits]

yxsamliu wrote:

> > > @YonahGoldberg gentle ping — the regression is fixed and the new benchmark run looks clean. Any further concerns, or is this good to go?
> > 
> > 
> > Do we not care about the regression in neatosplines.ll? Looks like there's some extra unnecessary vectorization that happens.
> 
> Got latest up-to-date result: [dtcxzyw/llvm-opt-benchmark-nightly#64](https://github.com/dtcxzyw/llvm-opt-benchmark-nightly/pull/64)

we got rid of the regression but also lost some optimizations. I will try recovering them.

https://github.com/llvm/llvm-project/pull/165159