[llvm] [NVPTX] deprecate nvvm.rotate.* intrinsics, cleanup funnel-shift handling (PR #107655)

[Artem Belevich via llvm-commits]

Artem-B wrote:




> the way we were lowering 64 bit rotation isn't actually correct.
> 
> Here's an alive proof in llvm that shows the expansion is incorrect https://alive2.llvm.org/ce/z/RQVDDG

Are you saying that your patch produces these invalid results? Or that LLVM without your patch miscompiles?

> I've move to just using the SelectionDAG expansion for all rotate instructions, this will properly convert of a 32-bit funnel shift when appropriate and expand to a valid multi-instruction expression otherwise.
> Could you please take another look?

I do not see any further changes on the pull requests since I've stamped it last week. Is the issue fixed the this pull request? Or do you want me to help figuring out where things still go wrong?


https://github.com/llvm/llvm-project/pull/107655