[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:

> > > Seems like we are still waiting for confirmation on this one?
> > 
> > 
> > My understanding from @mikulas-patocka in #106846 is that there was no regression on current main after merging over a week ago, so I think we should be good to go.
> > Not sure what the remaining time-line is, but it would be good to make sure the regression is fixed in the release.
> 
> There was a regression - the pointless move of %r14 to %r15 and back and loading %rsi from the stack that I mentioned in #106846 is actually a regression that was recently added to git. With older clang-22 from Debian Sid (version 1:22~++20250731080150+be449d6b6587-1~exp1), these pointless instructions are not generated.

Right, but was there a runtime regression? after all, without the patch, the Python on ARM64 macos regresses by 2-3% for end-to-end workloads

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