[llvm] [LoopUnroll] Cap unrolling and peeling in very large loop nests (PR #221702)

[Madhur Amilkanthwar via llvm-commits]

madhur13490 wrote:

> Before we move forward with this change, could you test out [2312a06](https://github.com/llvm/llvm-project/commit/2312a06c87207d28c83707db019b4053ceea9bae)'s `ForgetAllSCEV`? It seems to target a similar use case, but instead of calling `forgetTopmostLoop()`, it uses `forgetAllLoops()`.

I prototyped this but unfortunately this does not benefit as much as disabling. It gave just 8% speedup on roms (vs. 35%) benchmark. I would like to push this ahead because returns on this investment are quite good and promising. 

CC @alinas 


https://github.com/llvm/llvm-project/pull/221702