[llvm] [AArch64] Enable CmpBcc fusion for Neoverse-v2 (PR #90608)

[Sjoerd Meijer via llvm-commits]

sjoerdmeijer wrote:

> > I actually suggested that we remove this from some cores recently, as when I've tried it in the past (a long time ago now) the performance hasn't looked great. Perhaps things have changed since then. It's certainly true that the cores can do this fusion, my suspicion was that the implementation was a little too aggressive in llvm and led to worse codegen in places from knock-on effects.
> > Can you give more details on the "benchmark that improves significantly"? Do you happen to know why it improves?
> 
> The benchmark is Eigen, and we've seen 15% improvement. There are plenty of
> 
> ```
> cmp	
> str
> b.cc/b.eq/b.lt
> ```
> 
> code sequences there

I agree with Dave that we need to look into this a little bit more and double check that this codegen change is actually responsible for the performance uplift (and not e.g. because of some secondary effects). 

https://github.com/llvm/llvm-project/pull/90608