[Mlir-commits] [llvm] [mlir] [XeGPU][Transform] Add XeGPU array length optimization pass (PR #194062)

[Jianhui Li]

================
@@ -118,5 +118,20 @@ def XeGPUSgToWiDistributeExperimental : Pass<"xegpu-sg-to-wi-distribute-experime
                            "vector::VectorDialect", "index::IndexDialect"];
 }
 
+def XeGPUArrayLengthOptimization : Pass<"xegpu-array-length-optimization"> {
+  let summary = "Optimize XeGPU ops by introducing array_length attribute";
----------------
Jianhui-Li wrote:

I would say "Optimize XeGPU ops via 2D Block Array Load". 

For the optimization pass/phase name - I think "2D Block Array Load" may read better for external folks. Array length doesn't link to 2d block load hardware feature, so hard to understand if not familiar with the term. 

https://github.com/llvm/llvm-project/pull/194062