[Mlir-commits] [mlir] [mlir][nvgpu] Rename `mbarrier.test.wait` to `mbarrier.test`. (PR #96505)

[Guray Ozen]

grypp wrote:

We are not bound with PTX naming in `nvgpu` dialect. But this PR changes the name of the existing OP without a clear motivation.  

The `nvgpu` dialect serves as a bridge the `nvvm` dialect and higher-level dialects such as `memref` or `vector`. While the `nvvm` dialect is very close to PTX, it still abstracts some elements at a higher level than raw PTX. It's more like 1:N mapping. 

The high-level bit in nvgpu's `mbarrier` OP family is its ability to encapsulate multiple memory barriers into a single SSA value. These barriers can then be indexed and accessed using another SSA value. This level of abstraction and dynamic interaction is not feasible within the `nvvm` dialect due to its reliance on low-level.

https://github.com/llvm/llvm-project/pull/96505