[llvm] [AArch64] Add flag to control unrolling for small multi-exit loops (PR #131998)

[Florian Hahn via llvm-commits]

https://github.com/fhahn commented:

> > I am not sure about adding a very specialized option for this. Is there a reason we couldnt take the right decision per uarch to start with? It sounds like that is planned down the road already?
> 
> Well I was thinking that someone might be targeting a generic CPU with a particular architecture (perhaps this happens when targeting Android?), but still want a way to get the benefit of unrolling multi-exit loops such as std::find. Having said that, if the consensus is to avoid introducing a flag I can remove it?

I don't think we generally encourage users to use internal LLVM options, they mostly for LLVM developers, unless expose explicitly via `Clang` or other drivers. And those flags are also not guaranteed to be stable.

Another alternative as middle ground would be to introduce a target feature instead of a flag, this can be both specified via the command line and it will be easy to add it for CPUs where it's profitable.

But I think it would still be good to enable it for the CPUs that are known to benefit straigth away.

> Having the option sounds useful for debugging - we can use it to disable the unrolling if we find issues. Perhaps rename it to -aarch64-small-multi-exit-loop-unroll-factor?
> 
> I gave this a try and it seemed OK. The code size, which was the thing I was worried about, did not increase by too much.

If it is just about disabling unrolling for such loops, it might be good to check if the existing `unroll-runtime-multi-exit` option is explicitly set to false? This is something the current implementation for Apple CPUs doesn't do, but I think it would be reasonable to expect this flag to override the target specific decisions?

https://github.com/llvm/llvm-project/pull/131998