[llvm] [AMDGPU] (x or y) xor -1 -> x nor y (PR #130264)

[Nicolai Hähnle via llvm-commits]

nhaehnle wrote:

> > The problem I have here is that if I am using ballot
> 
> Ballot is not core to the pattern here, the ballot tests should be separate. But also the wave64 ballot should work on wave32

Looking at this with Ana offline, it seems that inverse.ballot.i64 doesn't work on wave32 yet. That should be fixed first.

As for the rest, I think the use of ballot/inverse.ballot here is very reasonable. Our calling conventions generally pass i1 as 32-bit values, and so without ballot/inverse.ballot the boolean logic tends to get folded into the v_cndmask/v_cmp that is generated as part of the calling convention lowering, which makes it difficult to test this particular pattern. Ballot/inverse.ballot is the most natural way to express what we want to do here.

https://github.com/llvm/llvm-project/pull/130264