[Mlir-commits] [mlir] [mlir] Added `Convergent` trait that matches LLVM's semantics (PR #152358)

[Nikolay Panchenko]

npanchen wrote:

> @npanchen Do you need this trait only for nvvm.barrier0? This is a special op that every thread in a CTA must run, so it’s convergent. Do you have another op that you think also requires it in your case?

I cannot definitively answer yes or no. At least my current workaround, to propagate convergency through a call graph, only relies on `nvvm.barrier0`. We do use some other operations I've previously marked as convergent, but I don't see any runtime issues so far.
fyi: `nvvm.barrier0` problem was uncovered recently due to recent changes in JT

https://github.com/llvm/llvm-project/pull/152358