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

[Md Abdullah Shahneous Bari]

================
@@ -556,11 +556,20 @@ LogicalResult LoadNdOp::verify() {
     }
   }
 
+  // Handle array_length. Two result shape conventions are accepted:
+  //   * Legacy: leading array_length dimension prepended, e.g. descriptor
+  //     16x16 with array_length=2 -> [2, 16, 16].
+  //   * Stacked 2D: array blocks stacked along the non-FCD (first) dimension,
+  //     e.g. descriptor 16x16 with array_length=2 -> [32, 16].
   auto array_len = tdescTy.getArrayLength();
-  if (array_len > 1)
-    tdescShape.insert(tdescShape.begin(), array_len);
+  SmallVector<int64_t> stackedShape(tdescShape);
----------------
mshahneo wrote:

Changed the var names.

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