[Mlir-commits] [mlir] [MLIR][XeGPU] Add blocking and subgroup to lane distribution support for ConvertLayout operation (PR #183837)

[Artem Kroviakov]

================
@@ -260,7 +248,17 @@ bool XeGPUBlockingPass::needsUnroll(Operation *op) const {
         std::optional<SmallVector<int64_t>> tileShape = getTileShape(result);
         return tileShape.has_value() && isUnrollable(result, *tileShape);
       });
-  return hasUnrollableOperands || hasUnrollableResults;
+  // ConvertLayoutOp must be processed to drop the inst_data in the layout
+  bool isConvertLayoutWithInstData = false;
+  if (isa<xegpu::ConvertLayoutOp>(op)) {
----------------
akroviakov wrote:

```suggestion
  if (auto convertLayoutOp = dyn_cast<xegpu::ConvertLayoutOp>(op)) {
```

https://github.com/llvm/llvm-project/pull/183837