[mlir][xegpu] Add OptimizeBlockLoads pass.

mlir/include/mlir/Dialect/XeGPU/Transforms/Passes.td

@@ -80,4 +80,14 @@ def XeGPUVectorLinearize : Pass<"xegpu-vector-linearize"> {
                            "scf::SCFDialect", "ub::UBDialect", "vector::VectorDialect"];
 }
 
+def XeGPUOptimizeTranspose : Pass<"xegpu-optimize-transpose"> {
+  let summary = "Optimize XeGPU loadNd operations feeding into vector.transpose";
+  let description = [{
+    This pass rewrites XeGPU loadNd operations that feed into vector.transpose
+    into more optimal forms to improve performance.
+  }];
+  let dependentDialects = ["memref::MemRefDialect", "xegpu::XeGPUDialect",
+                           "vector::VectorDialect"];
+}
+
 #endif // MLIR_DIALECT_XEGPU_TRANSFORMS_PASSES_TD

mlir/include/mlir/Dialect/XeGPU/Transforms/Transforms.h

@@ -61,7 +61,8 @@ struct UnrollOptions {
 
 /// Appends patterns for folding aliasing ops into XeGPU ops into `patterns`.
 void populateXeGPUFoldAliasOpsPatterns(RewritePatternSet &patterns);
-
+/// Appends patterns for optimizing transpose operations into `patterns`.
+void populateXeGPUOptimizeTransposePatterns(RewritePatternSet &patterns);
 /// Appends patterns for XeGPU SIMT distribution into `patterns`.
 void populateXeGPUSubgroupDistributePatterns(RewritePatternSet &patterns);
 /// Appends patterns for moving function body into gpu.warp_execute_on_lane0 op.

mlir/lib/Dialect/XeGPU/Transforms/CMakeLists.txt

@@ -6,6 +6,7 @@ add_mlir_dialect_library(MLIRXeGPUTransforms
   XeGPUWgToSgDistribute.cpp
   XeGPUPropagateLayout.cpp
   XeGPUVectorLinearize.cpp
+  XeGPUOptimizeTranspose.cpp
 
   ADDITIONAL_HEADER_DIRS
   ${MLIR_MAIN_INCLUDE_DIR}/mlir/Dialect/XeGPU