[Mlir-commits] [mlir] [MLIR][XeGPU] Improve deinterleave/interleave/dpas_mx ops handling (PR #197223)

[Artem Kroviakov]

================
@@ -96,6 +96,8 @@ void buildGPUPassPipeline(OpPassManager &pm,
     pm.addNestedPass<gpu::GPUModuleOp>(createLoopInvariantCodeMotionPass());
     pm.addNestedPass<gpu::GPUModuleOp>(createCSEPass());
     pm.addNestedPass<gpu::GPUModuleOp>(xegpu::createXeGPUVectorLinearize());
+    pm.addNestedPass<gpu::GPUModuleOp>(createCanonicalizerPass());
----------------
akroviakov wrote:

The new ones are needed functionally. The `cse` and `canonicalize` before `LoopInvariantCodeMotionPass` can eliminate multiple thousands lines of code in certain examples. Not sure how much of a performance hit it is, though

https://github.com/llvm/llvm-project/pull/197223