[llvm] [TailDuplicator] Do not restrict the computed gotos (PR #114990)

[Florian Hahn via llvm-commits]

https://github.com/fhahn commented:

We found that perfoming aggressive tail-dup for blocks with computed gotos before register allocation is causing a 3% regression on some Python 3.12 workloads on macOS. This is due to missed optimizations after expanding the CFG too early.

I think it would be good to delay agressive tail-dup with computed gotos until register allocation, to allow more optimizations to trigger, as well as potentially saving some compile-time: https://github.com/llvm/llvm-project/pull/150911

https://github.com/llvm/llvm-project/pull/114990