[llvm] [GVN] Support rnflow pattern matching and transform (PR #162259)

[Madhur Amilkanthwar via llvm-commits]

madhur13490 wrote:

>Could you also please test this on the llvm test suite? This is the kind of change that has a high chance of introducing a hard to track down miscompile, unless tested well beforehand.
Also see above about having it under a flag at first, for similar reasons.

I tested llvm test-suite with the recent version, I don't see any crash or abnormal behaviour. 

About the flag: 
I don't mind having cl::opt flag but I prefer it to be on by default. I see your point about keeping the feature off by default, especially from a safety perspective. That said, having it off introduces an extra path that is rarely exercised and adds downstream operational overhead for us. I’d prefer to keep it on by default and rely on our existing monitoring/rollback mechanisms, so the behaviour we have is the one we actually test and maintain. 

I have tested this patch widely on various benchmarks and workloads and since the pattern is quite specific, I believe the miscompile is rare. 

IIUC, are you suggesting that we land this commit with off by default and then in the successive commit we just flip the switch to make the revert easy, in case of miscompiles?

https://github.com/llvm/llvm-project/pull/162259