[llvm] TableGen: Optimize super-register class computation (PR #134865)

[Björn Pettersson via llvm-commits]

bjope wrote:

> @nhaehnle , I'm trying to understand the impact of this for our downstream out-of-tree target. Some of our benchmarks has shown regressions, and one thing I've noticed is that the register pressure sets differ a lot. There is really no comments in the commit message about this really changing anything (except for register class names). And there aren't any test cases that are impacted.
> 
> (No idea yet if we get codegen diffs due to something special we have in our downstream fork.)
> 
> Just want to double check if this just is supposed to save compilation time, without any changes in codegen. Or am I really supposed to see impact in llc codegen with this patch?

I did check for some in-tree target and then the <target>GenRegisterInfo.inc files isn't impacted. So it looks like we have some downstream added property on the register classes that isn't inferred the same way after this patch. Hmm. We have probably 
been "lucky" with iteration order or something in the past. I'll keep digging on my end as I suspect this is a downstream issue.

https://github.com/llvm/llvm-project/pull/134865