[llvm] Use 256 bit register when AVX2 or higher is available. (PR #91721)

[via llvm-commits]

shamithoke wrote:

> > > > I can't see anything in Agner/instlatx64/uops.info to match this, and it'd be very unusual for a 256-bit instruction to be faster than the 128-bit variant.
> > > > Do you have a benchmark that I can run on my rocketlake machine?
> > > > @phoebewang Do you have any ideas what could be going on?
> > > 
> > > 
> > > I have a java test (that is taken through an internal JIT compiler) and I measured performance for this test. Can you please recommend what kind of benchmark can be built?
> > > I observe roughly 5% improvement while using AVX2.
> > 
> > 
> > I don't know what's the general benchmark used by community. I can try it with our internal benchmarks.
> 
> The result shows only two benchmarks are affected, which have both 0.15% improvement, hard to tell if they are noise.

Oh! That is very minimal change. In my case, the improvement is in 3-5% range. 
Thanks for running the benchmarks, @phoebewang 

https://github.com/llvm/llvm-project/pull/91721