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

[Paul Walker via llvm-commits]

paulwalker-arm wrote:

I don't believe we want or should be adding a new "public" way to control this optimisation and so we shouldn't add anything binding like a target feature flag.  The intent is really about having finer control for environments where it would not be easy to inject a new toolchain but a user is happy to try "secret" options using their existing toolchain, be it for functional or performance investigations.

I do agree we should be more aggressive in enabling this optimisation by default for the platforms we know it will benefit.

@fhahn: What is the down side to adding a dedicated option for this? If there's a problem with adding such flags then perhaps we should think about adding something akin to gcc's `-param` flag? Would there be benefit to having one flag with many properties verse having many flags?

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