[Mlir-commits] [mlir] [MLIR][GPU] Add xevm-attach-target transform pass. (PR #147372)

[Artem Kroviakov]

================
@@ -258,4 +258,38 @@ def GpuSPIRVAttachTarget: Pass<"spirv-attach-target", ""> {
   ];
 }
 
+def GpuXeVMAttachTarget : Pass<"xevm-attach-target", "mlir::gpu::GPUModuleOp"> {
----------------
akroviakov wrote:

This PR might be a good place to mention why it is better to relax the restriction and opt for more complexity in the pass implementation, apart from being stylistically aligned. For this specific pass and its nv/roc counterparts, there is no cross-module logic apparent.
If one wants to run a pass on a specific op, why can't/shouldn't one rely on the existing MLIR op-specific pass infrastructure and instead should replicate the traversal logic explicitly?
Is it performance related in some significant way?
If there are corner cases, shouldn't the existing test suite include them to highlight why the op-specific pass must be avoided?

https://github.com/llvm/llvm-project/pull/147372