[llvm-branch-commits] [llvm] release/21.x: [TailDup] Delay aggressive computed-goto taildup to after RegAlloc. (#150911) (PR #151680)

[Florian Hahn via llvm-branch-commits]

fhahn wrote:

> There was 9% runtime regression on a loop that just adds numbers (so that it stresses this code path significantly). There was no measurable regression on Ajla self-compilation (the difference is just 0.8% - which may be jitter).


Thanks for clarifying! The regression seems like an X86 backend issue, unrelated directly to tail duplication for computed gotos. Given that there is no (or a very small) end-to-end regression, I think the change should be included in the release, or an argument should be made for reverting/adjusting https://github.com/llvm/llvm-project/pull/150911 (the version of the PR landed on `main`)

https://github.com/llvm/llvm-project/pull/151680