[Mlir-commits] [mlir] feat(linalg): enable lowering/decomposing scalable pack ops (PR #200216)

[Artem Gindinson]

================
@@ -249,6 +243,14 @@ FailureOr<LowerPackResult> linalg::lowerPack(RewriterBase &rewriter,
   SmallVector<int64_t> stripMinedShape(packedTensorType.getShape());
   applyPermutationToVector(stripMinedShape, packedToStripMinedShapePerm);
 
+  // Also compute the mixed (static+dynamic) strip-mined sizes for the
+  // expand_shape output. This is needed to support dynamic inner tile sizes,
+  // since the 3-arg ExpandShapeOp builder cannot infer shapes when multiple
+  // dynamic dims appear in a single reassociation group.
----------------
AGindinson wrote:

Nit / call for a higher-level comment:
```suggestion
  // since the shapes cannot be inferred automatically when multiple dynamic dims
  // appear in a single reassociation group during ExpandShapeOp construction.
```

https://github.com/llvm/llvm-project/pull/200216