[llvm-branch-commits] [llvm] [AMDGPU] Enable atomic optimizer for divergent i64 and double values (PR #96934)

[Vikram Hegde via llvm-branch-commits]

vikramRH wrote:

> > > > [AMDGPU] Enable atomic optimizer for divergent i64 and double values
> > > 
> > > 
> > > Needs some i64 tests
> > 
> > 
> > added new i64 tests, however I see there currently exists an issue with DPP path where dpp combine partially fuses the mov_dpp pieces causing machine CSE crash. I have proposed #97413 for now. what would be the correct way forward here ?
> 
> You didn't include a (very necessary) test in #97413, but DPP instructions shouldn't be candidates for trivial CSE in the first place?

sorry about that, just wanted to bring this up first (I will update the PR with a test). The issue is not with DPP instructions themselves but with the REG_SEQUENCE which is generated after fusing the 32 bit pieces.

https://github.com/llvm/llvm-project/pull/96934