[llvm] [AArch64][InstCombine] Eliminate redundant barrier intrinsics (PR #112023)

[Danila Malyutin via llvm-commits]

danilaml wrote:

@david-arm This change mostly parallels optimizations done for fences in `InstCombinerImpl::visitFenceInst` and at least there it was beneficial. It's hard for me to come up with a situation where `dmb; some instrs; dmb` would be faster than `some instrs; dmb` except some really weird hw quirk not sure we could account for in IR anyway (and again, fences would be affected in the same way).

https://github.com/llvm/llvm-project/pull/112023